完成微盘入退场历史分位多日平滑实验

This commit is contained in:
baobao 2026-10-28 13:07:24 +08:00
parent c51fb7fbd4
commit c57ed900e2
23 changed files with 10301 additions and 0 deletions

View File

@ -2416,3 +2416,18 @@ QMT仍有序号缺口,16:23确认1441965、最高1442003、暂存32条。16:24
真机判收:50只名单与60条历史就绪,页面8%/85%/50正确;最新分位8.33%,三项触发均false。新规则盘中未冻结,旧冻结文件摘要不变。实际连续确认继续推进,无新增验签拒绝或重连,对账与现金差异0。三宽度浏览器复现通过。开发机及155断网全量回归通过,微盘55项、监控12项通过。
还欠着:新版本完整自然入退场和成交尚未判收。10月29日08:50按既有流程冻结,09:30至10:00按信号执行;不增加观察和纸面等待。日常复核任务已改为50股规则。没有改原策略持仓或在途委托。
## 2026-10-28 · 微盘轮动入退场分位的多日平滑实验
做了什么:以已部署8%入场、85%退出、50只为基线,冻结3/5/10/20交易日均值和中位数,分别只平滑入场、只平滑退出、两侧平滑,共24种。只平滑每日历史百分位,成长瓦解和账户方法不变,使用2024年前数据预热。另做单日分位桥接,678天信号及账户与基线一致。
动了哪些文件:新增research/micro_smoothing_20261028中的协议、离线实验、测试、独立复算、报告、完整结果与两张图。原标定脚本和线上代码没有修改。
结果:2024-01-02至2026-10-23基线累计79.44%、最大回撤21.07%。双侧3日均值51.69%,3日中位数61.97%,最大回撤均未改善。双侧10/20日平滑漏掉2026年初高位退出,回撤扩大到31.73%。24种平滑在30与60基点费用下收益均低于基线。结论仅限固定阈值下的历史比较,宏观轮次稀少,不声称统计显著。
部署方式:未部署,155与188未访问或修改,现有8%/85%/50只自动运行规则保持。研究文件经Git保存,不把离线实验自动升级为生产策略。
判收:开发机断网隔离9项测试通过;独立审核27个账户、52次前缀重跑、17628条指标,最大现金净值误差4.22e-15。两张图已检查。线上交易效果未判收。完整报告为research/micro_smoothing_20261028/微盘轮动多日平滑实验.md。
还欠着:真实策略自然轮次仍需模拟运行。历史退市与简称覆盖、复权单位和开盘成交假设沿用前次局限。本实验不要求再调参或部署。

View File

@ -0,0 +1,50 @@
"""复核全部主账户、滚动窗口与历史前缀,独立计算余额和估值。"""
import sys
import numpy as np
from pathlib import Path
sys.path.insert(0,str(Path(__file__).resolve().parent))
from experiment import ROOT,OUT,BASE,read,load_data,candidates,signal_series,run,choose,save,digest,no_network
from audit import audit_account
def main():
sys.addaudithook(no_network)
p,data,obs,coverage=load_data();cands=candidates(p)
bridge={'id':'percentile_1','method':'mean','window':1,'scope':'both'};allspecs=cands+[bridge]
accounts=read(OUT/'accounts_30.json.gz');saved=read(OUT/'signals.json.gz')['rows'];grid=read(OUT/'summary_30.json');stress=read(OUT/'summary_60.json')
ids={r['date'][:10]:i for i,r in enumerate(obs)};dates=[r['date'] for r in data]
independent={};prefixes=[];signal_count=0
for c in allspecs:
name=c['id'];ss=signal_series(data,obs,c)
independent[name]=audit_account(data,accounts[name],(8,85,50))
assert grid[name]['total_return']>=stress[name]['total_return']
assert all(s['growth_break']==row['signal']['growth_break'] for s,row in zip(ss,data))
for d,r in zip(data,ss):
i=ids[d['date']];values=[o['activity_percentile'] for o in obs[i-c['window']+1:i+1]]
q=float(np.median(values) if c['method']=='median' else np.mean(values))
assert abs(r['filtered_percentile']-q)<1e-12
if c['scope'] in ('entry','both'):assert r['low']==(q<.08),(name,d['date'],'entry')
if c['scope'] in ('exit','both'):assert r['high']==(q>.85),(name,d['date'],'exit')
signal_count+=1
for year in ('2024','2025'):
n=sum(d[:4]<=year for d in dates);cut=data[n-1]['date']
truncated_obs=[r for r in obs if r['date'][:10]<=cut]
truncated=signal_series(data[:n],truncated_obs,c)
assert truncated==ss[:n]
account=run(data[:n],truncated)
assert account.basket.records==accounts[name]['daily'][:n]
prefixes.append({'method_id':name,'through':cut,'days':n,'max_difference':0.})
sequential=read(OUT/'sequential.json.gz');navs={k:[r['nav'] for r in a['daily']] for k,a in accounts.items()}
for selected in sequential['selection']:
n=sum(d<=selected['train_end'] for d in dates)
winner,_=choose(cands,{k:v[:n] for k,v in navs.items()},dates[:n],n)
assert winner==selected['method_id']
independent['sequential']=audit_account(data,sequential['accounts']['30'],(8,85,50))
assert accounts['baseline']==accounts['percentile_1']
out={'independent_accounts':independent,'prefix_checks':prefixes,'signal_rows_checked':signal_count,'growth_unchanged':True,
'percentile_bridge_exact_account_match':True,'higher_cost_never_better':True,'sequential_selection_uses_training_only':True,
'note':'独立余额、逐笔先卖后买、交易约束与估值复算;所有26条账户在2024和2025末截断重跑。未验证历史数据库缺失的股票。'}
save('audit.json',out);print('审核通过:',len(independent),'账户,',len(prefixes),'截断重跑,',signal_count,'条指标;现金净值最大误差',max(v['max_cash_nav_error'] for v in independent.values()))
if __name__=='__main__':main()

View File

@ -0,0 +1,30 @@
{
"sha256": {
"audit_smoothing.py": "bfdb190892899f5c0047ca388adfa552f431e0fb5264561d9db494d8c89fb101",
"experiment.py": "c65b334a559baf93475aa8f7aede56984718239d9d05aa12e239ee62f4de8967",
"figures/净值与退出延迟.png": "3e21c2afbb428ee85b2479c4fc93b6724700299120eae22c0f3531b905380fda",
"figures/均值中位数对照.png": "8dd37a14b32ad0711dd0198381fc04b03d75d05c62fc29b8ce3ed2569b8c4c66",
"protocol.json": "cec72a585f671b78134d24eb993dd707deb67849376af673f8937f3314e85a8b",
"protocol.sha256": "3d51b24866a2397569bb1673698eff3731af0ffb6e73a8591fb0f7853c96a4c2",
"report.py": "d802acced9118c061cfcf79ca6ed623aa3932b077dd35e680f1b6a98806a6e74",
"results/accounts_30.json.gz": "14adebabf85d784688eb42369b4ef3544582f849561568c1c628002a5b187ecb",
"results/audit.json": "a9d26345b3b2589aba8663193060d9c32a4bd8e0868d86fdb6d2f5b3e5d624a5",
"results/comparison.csv": "0d799460676f01606777605b31ea60f81e847aa5fbb377a74cbb52e7c2e20626",
"results/coverage.json": "9569d0b160037bbab8fad87a2821b32e18222df0a78ccd24c744579062c65cbf",
"results/cycles_comparison.json": "b74aa05fab7fe8224dd321516416f7cb53a465ce9fff54df3e905ba1fc580124",
"results/interpretation.json": "b9f7bc767596ede7831d86db211c5445313680100f846dd686187787ffef097d",
"results/manifest.json": "dac98453e6c6afa40f6a859cf1339a33e89cc9a767e2a9b0f7f3cf2d533d9188",
"results/navs.npz": "172525feb8d1535c7e2fab2aaa5029bb8c9617e87378be0e1a43a6b65fa8d67b",
"results/sequential.json.gz": "c9e8094042062066f5a2f5d53c286d2fadce0c8ed8ade125ff0dc82ae0a004c6",
"results/signals.json.gz": "4e5d36a136eebff2b114b97be01e6babe24fcac1239f002751a7505abb415f02",
"results/summary_30.json": "1fa37823db502fc37bf0a59f9cbc7988ca1ff36490e0908ef616744645bdc8f0",
"results/summary_60.json": "3aca7477e64911c9c4b2a0748ac681d2e95fd5f513687ca8ed9698ddeb78be33",
"test_smoothing.py": "114cad833e93614cbe5736d26e3b88729da41f24aa4375a33e5ae2ad627f3ba4",
"微盘轮动多日平滑实验.md": "beb515287000257bed69a010224ee38d47f9ec2047715c561989aa2fe0985e12"
},
"environment": "开发机,Python 3.12.14,/tmp/micro-calibration-venv;研究过程禁联网",
"unit_test_result": "9 tests OK",
"independent_audit": "27 accounts / 52 prefixes / 17628 signal rows passed",
"production_changes": false,
"repeatability": "统一CSV换行后重跑,两个费用的摘要、全部账户、连续账户与信号文件摘要完全相同"
}

View File

@ -0,0 +1,191 @@
"""百分位多日平滑实验。复用冻结离线账户,所有业务数据仅从本地读取。"""
import bisect
import csv
import datetime as dt
import gzip
import hashlib
import json
import math
import statistics
import sys
from pathlib import Path
import numpy as np
ROOT=Path(__file__).resolve().parent
BASE=ROOT.parent/'micro_calibration_20261028'
sys.path.insert(0,str(BASE))
from prepare import read, no_network, threshold_rows
from calibrate import MeasuredBasket, metric, summary
from engine import TimingStrategy
OUT=ROOT/'results'
def digest(path):return hashlib.sha256(path.read_bytes()).hexdigest()
def save(name,value):
p=OUT/name;p.parent.mkdir(exist_ok=True,parents=True)
text=json.dumps(value,ensure_ascii=False,allow_nan=False,indent=2)+'\n'
if name.endswith('.gz'):p.write_bytes(gzip.compress(text.encode(),mtime=0))
else:p.write_text(text)
def smooth(values,window,method):
if not isinstance(window,int) or window<1 or method not in ('mean','median'):raise ValueError('非法平滑配置')
result=[]
for i in range(len(values)):
part=values[max(0,i-window+1):i+1]
valid=len(part)==window and all(isinstance(x,(int,float)) and math.isfinite(x) and 0<=x<=1 for x in part)
result.append((statistics.mean(part) if method=='mean' else statistics.median(part)) if valid else None)
return result
def candidates(p):
return [{'id':'baseline','method':'raw','window':1,'scope':'none'}]+[
{'id':f'{m}_{w}_{s}','method':m,'window':w,'scope':s}
for m in p['methods'] for w in p['windows'] for s in p['scopes']]
def load_data():
p=read(ROOT/'protocol.json')
assert digest(ROOT/'protocol.json')==(ROOT/'protocol.sha256').read_text().split()[0]
coverage=read(BASE/'coverage.json');cache=BASE/'cache/daily.json.gz'
assert digest(cache)==coverage['cache_sha256']
obsfile=ROOT.parent/'macro_timing_20261026/results/rebalancing_observation/observations.json.gz'
assert digest(obsfile)==coverage['source_hashes']['results/rebalancing_observation/observations.json.gz']
data=read(cache);obs=read(obsfile)['groups']['small']
signals=threshold_rows([r['signal'] for r in data],obs,[5,8,85,95])
for r,s in zip(data,signals):r['signal']=s
dates=[o['date'][:10] for o in obs];assert dates==sorted(set(dates))
indices={d:i for i,d in enumerate(dates)};first=indices[data[0]['date']]
assert first>=max(p['windows'])-1
for r in data:
o=obs[indices[r['date']]];assert abs(o['activity_percentile']-r['signal']['activity_percentile'])<1e-12
assert len(r['members'])==100 and len(set(r['members'][:50]))==50
assert len(data)==678 and [data[0]['date'],data[-1]['date']]==p['sample']
metadata={'start':data[0]['date'],'end':data[-1]['date'],'days':len(data),'warmup_start':dates[first-max(p['windows'])+1],
'warmup_observations':max(p['windows'])-1,'cache_sha256':digest(cache),'observations_sha256':digest(obsfile),
'engine_sha256':digest(BASE/'engine.py'),'calibrate_sha256':digest(BASE/'calibrate.py')}
return p,data,obs,metadata
def signal_series(data,obs,spec):
by={r['date'][:10]:i for i,r in enumerate(obs)}
values=[r['activity_percentile'] for r in obs]
filtered=smooth(values,spec['window'],spec['method']) if spec['method']!='raw' else values
result=[]
for row in data:
s=row['signal'];q=filtered[by[row['date']]]
if q is None:raise ValueError('完整历史窗口不足,不产生交易信号')
low=s['activity']<s['thresholds']['8'];high=s['activity']>s['thresholds']['85']
if spec['scope'] in ('entry','both'):low=q<.08
if spec['scope'] in ('exit','both'):high=q>.85
result.append({**s,'low':low,'high':high,'filtered_percentile':q,
'entry_reading':q if spec['scope'] in ('entry','both') else s['activity_percentile'],
'exit_reading':q if spec['scope'] in ('exit','both') else s['activity_percentile']})
return result
def run(data,signals,bps=30,switches=None):
s=TimingStrategy(cost=bps/10000);s.basket=MeasuredBasket(cost=bps/10000)
current='baseline' if isinstance(signals,dict) else None
for i,row in enumerate(data):
if switches and row['date'] in switches:current=switches[row['date']]
sig=signals[current][i] if current else signals[i]
s.step(row['date'],row['members'][:50],row['rows'],row['limits'],set(row['risk']),sig)
s.decisions[-1].update(filtered_percentile=sig['filtered_percentile'],entry_reading=sig['entry_reading'],exit_reading=sig['exit_reading'],method_id=current)
return s
def details(s):
return {'daily':s.basket.records,'trades':s.basket.trades,'decisions':s.decisions,'failures':s.basket.failures}
def signal_stats(signals):
out={}
for k in ('low','high','growth_break'):
flags=[r[k] for r in signals];out[k+'_days']=sum(flags)
out[k+'_episodes']=sum(v and (i==0 or not flags[i-1]) for i,v in enumerate(flags))
return out
def cycle_comparison(base,candidate,dates):
# 按实际持仓日期重叠匹配,不把不同轮次强行解释成固定延迟。
pos={d:i for i,d in enumerate(dates)};matches=[];used=set()
def interval(c):return pos[c['entry_date']],pos[c['exit_date']] if c['exit_date'] else len(dates)-1
for b in base:
bi,bj=interval(b);overlaps=[]
for j,c in enumerate(candidate):
if j in used:continue
ci,cj=interval(c);overlap=max(0,min(bj,cj)-max(bi,ci)+1)
if overlap:overlaps.append((overlap,-abs(ci-bi),j))
if not overlaps:
matches.append({'baseline_id':b['id'],'candidate_id':None,'status':'missing_cycle'});continue
j=max(overlaps)[2];used.add(j);c=candidate[j]
matches.append({'baseline_id':b['id'],'candidate_id':c['id'],'status':'overlap_match',
'baseline_entry':b['entry_date'],'candidate_entry':c['entry_date'],
'entry_delay_trade_days':pos[c['entry_date']]-pos[b['entry_date']],
'same_entry_signal':b['signal_date']==c['signal_date'],
'baseline_exit':b['exit_date'],'candidate_exit':c['exit_date'],
'exit_delay_trade_days':pos[c['exit_date']]-pos[b['exit_date']] if c['exit_date'] and b['exit_date'] else None,
'baseline_return':b['return_value'],'candidate_return':c['return_value']})
for j,c in enumerate(candidate):
if j not in used:matches.append({'baseline_id':None,'candidate_id':c['id'],'status':'added_cycle','entry_date':c['entry_date'],'exit_date':c['exit_date']})
return matches
def choose(cands,navs,dates,n):
metrics={c['id']:metric(navs[c['id']][:n],dates[:n]) for c in cands}
order=sorted(cands,key=lambda c:(-metrics[c['id']]['utility'],c['id']!='baseline',c['window'],c['method'],c['scope']))
return order[0]['id'],metrics[order[0]['id']]
def main():
sys.addaudithook(no_network)
p,data,obs,coverage=load_data();OUT.mkdir(exist_ok=True)
dates=[r['date'] for r in data];cands=candidates(p);assert len(cands)==p['candidate_count']
bridge={'id':'percentile_1','method':'mean','window':1,'scope':'both'}
allspecs=cands+[bridge];signals={c['id']:signal_series(data,obs,c) for c in allspecs}
outputs={};navs={};all_details={};bridge_diff=[]
for i,(a,b) in enumerate(zip(signals['baseline'],signals['percentile_1'])):
if (a['low'],a['high'])!=(b['low'],b['high']):bridge_diff.append({'date':dates[i],'baseline_low':a['low'],'baseline_high':a['high'],'percentile_low':b['low'],'percentile_high':b['high'],'percentile':b['activity_percentile']})
for bps in p['cost_bps']:
results={}
for c in allspecs:
name=c['id'];s=run(data,signals[name],bps);m=summary(s,data,(8,85,50),bps)
m.update(spec=c,signals=signal_stats(signals[name]),entry_vetoes=sum(d['entry_vetoed_by_high'] for d in s.decisions))
results[name]=m
if bps==30:
navs[name]=[r['nav'] for r in s.basket.records];all_details[name]=details(s)
outputs[str(bps)]=results
save(f'summary_{bps}.json',results);print('完成成本',bps,'组合',len(results),flush=True)
baseline=outputs['30']['baseline']
assert abs(baseline['total_return']-.7944394273468172)<1e-10
assert abs(baseline['max_drawdown']-(-.21073920900975784))<1e-10
save('accounts_30.json.gz',all_details)
comparison={c['id']:cycle_comparison(baseline['cycles'],outputs['30'][c['id']]['cycles'],dates) for c in cands}
save('cycles_comparison.json',comparison)
save('signals.json.gz',{'specs':allspecs,'rows':{name:[{k:r[k] for k in ('date','activity_percentile','filtered_percentile','entry_reading','exit_reading','low','high','growth_break')} for r in rows] for name,rows in signals.items()},'bridge_differences':bridge_diff})
switches={};selection=[]
for year in ('2024','2025'):
n=sum(d[:4]<=year for d in dates);winner,train=choose(cands,navs,dates,n);switches[dates[n-1]]=winner
selection.append({'train_end':dates[n-1],'apply_year':str(int(year)+1),'method_id':winner,'train':train})
rolling={}
for bps in p['cost_bps']:
s=run(data,signals,bps,switches);rolling[str(bps)]={'summary':summary(s,data,(8,85,50),bps),**details(s)}
save('sequential.json.gz',{'selection':selection,'accounts':rolling})
with (OUT/'comparison.csv').open('w',newline='') as f:
w=csv.writer(f,lineterminator='\n');w.writerow(['method_id','method','window','scope','cost_bps','total_return','annualized','drawdown','2024','2025','2026','cycles','closed_cycles','trades','fees_initial_capital','mean_turnover','exposure_days','low_signal_days','high_signal_days','entry_vetoes'])
for bps,results in outputs.items():
for name,m in results.items():
c=m['spec'];w.writerow([name,c['method'],c['window'],c['scope'],bps,m['total_return'],m['annualized_return'],m['max_drawdown']]+[m['annual'][y]['return'] for y in ('2024','2025','2026')]+[m['cycle_count'],m['closed_cycles'],m['trade_count'],m['fees_initial_capital'],m['mean_turnover'],m['exposure_days'],m['signals']['low_days'],m['signals']['high_days'],m['entry_vetoes']])
np.savez_compressed(OUT/'navs.npz',dates=np.array(dates),ids=np.array(list(navs)),nav=np.array(list(navs.values())))
save('coverage.json',coverage)
save('manifest.json',{'protocol_sha256':digest(ROOT/'protocol.json'),'experiment_sha256':digest(Path(__file__)),**coverage,'completed_at':dt.datetime.now().astimezone().isoformat(),'accounts':len(allspecs)*2,'unique_primary_paths':len({hashlib.sha256(np.asarray(v,dtype='<f8').tobytes()).hexdigest() for k,v in navs.items() if k!='percentile_1'}),'bridge_differences':bridge_diff,'sequential_selection':selection})
print('基线',baseline['total_return'],baseline['max_drawdown'],'边界差异',len(bridge_diff),flush=True)
for c in cands:
if c['scope'] in ('both','none'):
m=outputs['30'][c['id']];print(c['id'],round(m['total_return']*100,2),round(m['max_drawdown']*100,2),m['cycle_count'],m['closed_cycles'],flush=True)
if __name__=='__main__':main()

Binary file not shown.

After

Width:  |  Height:  |  Size: 280 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 137 KiB

View File

@ -0,0 +1,56 @@
{
"version": "micro-percentile-smoothing-v1-20261028",
"frozen_at": "2026-10-28T12:57:47.406640+08:00",
"baseline": [
8,
85,
50
],
"sample": [
"2024-01-02",
"2026-10-23"
],
"methods": [
"mean",
"median"
],
"windows": [
3,
5,
10,
20
],
"scopes": [
"entry",
"exit",
"both"
],
"candidate_count": 25,
"cost_bps": [
30,
60
],
"meaning": "平滑已计算的每日历史百分位,窗口含信号日及此前N-1个交易日。不是对8%或85%常数取平均,也不是先平滑原始成交额再重新排名。",
"warmup": "使用2024年前真实指标预热,要求完整窗口,不以未来或短窗口填补。所有历史分位均来自此前756个交易日,不含当日。",
"baseline_bridge": "基线沿用原始活跃度与线性插值8%/85%分位阈值严格比较;另列单日历史百分位8%/85%桥接对照,披露离散分位边界差异。",
"signals": "只入场:均值或中位数低于8%,退出沿用基线。只退出:入场沿用基线,高于85%退出。两者:两个分位均用同一窗口同一方法;成长瓦解定义不变,高位优先否决入场。",
"trading": "2024首日空仓,50股资格/市值排序和成员日换不变;收盘信号次日开盘,先卖后买,保留成员不重新等权;停牌/涨跌停/风险过滤及残余退出保持。",
"report": [
"全部25组及单日分位桥接对照",
"累计及逐年绝对收益、回撤、费用、换手、轮次",
"逐轮入退场日期、已结束轮收益与未结束估值",
"相同事件的触发延迟和因高位否决变化而新增或消失的轮次",
"30/60基点成本压力",
"按时间顺序选择的连续账户"
],
"sequential_diagnostic": "以2024已发生净值、2024至2025已发生净值分别选择下一年方法,评分沿用年化收益减回撤绝对值。平手优先基线,再较短窗口、方法和范围自然序。2024用基线启动,年末收盘换下一日信号,现金与持仓不重置。非独立留出,不据此自动改线上。",
"limits": [
"只有约三年且宏观独立轮次稀少,25组共享历史不是25份独立证据",
"旧历史已研究过,属于回溯实验,不声称未触碰样本外",
"可分割复权单位,单边30/60基点统一摩擦,不是实际整手含最低佣金账户",
"历史退市及简称覆盖不完整,沿用既有12条行情修复和估值规则",
"独立微盘绝对收益,不编造主观量化筹资延迟或两策略联合收益"
],
"operations": "开发机离线,禁止联网读取业务数据;不改155或188运行代码、参数或容器。",
"methodology_reference": "https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf"
}

View File

@ -0,0 +1 @@
cec72a585f671b78134d24eb993dd707deb67849376af673f8937f3314e85a8b protocol.json

View File

@ -0,0 +1,173 @@
"""读取既有实验结果生成图和中文报告,不接入线上。"""
import datetime as dt
import json
import hashlib
import statistics
from pathlib import Path
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from matplotlib import font_manager,ticker
from experiment import ROOT,OUT,read,no_network
import sys
sys.addaudithook(no_network)
FIG=ROOT/'figures';FIG.mkdir(exist_ok=True)
for path in ('/System/Library/Fonts/STHeiti Light.ttc','/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc'):
if Path(path).exists():
font_manager.fontManager.addfont(path);plt.rcParams['font.family']=font_manager.FontProperties(fname=path).get_name();break
plt.rcParams.update({'axes.unicode_minus':False,'font.size':11,'axes.spines.top':False,'axes.spines.right':False,'savefig.dpi':160,'figure.facecolor':'white','axes.titleweight':'bold'})
g=read(OUT/'summary_30.json');stress=read(OUT/'summary_60.json');p=read(ROOT/'protocol.json');audit=read(OUT/'audit.json')
accounts=read(OUT/'accounts_30.json.gz');signals=read(OUT/'signals.json.gz')['rows'];seq=read(OUT/'sequential.json.gz');manifest=read(OUT/'manifest.json')
def pct(x):return f'{x*100:+.2f}%'
def plain(x):return f'{x*100:.2f}%'
def name(k):
if k=='baseline':return '未平滑基线'
if k=='percentile_1':return '单日百分位对照'
m,w,scope=k.split('_');return f'{w}日'+('均值' if m=='mean' else '中位数')
def table(headers,rows):return '\n'.join(['| '+' | '.join(headers)+' |','| '+' | '.join(['---']*len(headers))+' |']+['| '+' | '.join(map(str,row))+' |' for row in rows])
both=['baseline']+[f'{m}_{w}_both' for m in p['methods'] for w in p['windows']]
fig,axes=plt.subplots(1,2,figsize=(12.5,6.1),layout='constrained')
labels=[name(k) for k in both];colors=['#637185']+['#3d7eff']*4+['#d8a632']*4
for ax,field,title in zip(axes,['total_return','max_drawdown'],['累计绝对收益','最大回撤幅度(越小越好)']):
vals=[g[k][field]*100*(1 if field=='total_return' else -1) for k in both]
bars=ax.barh(labels,vals,color=colors,height=.63)
ax.invert_yaxis();ax.set_xlim(0,max(vals)*1.17);ax.xaxis.set_major_formatter(ticker.PercentFormatter());ax.grid(axis='x',alpha=.18);ax.set_axisbelow(True);ax.set_title(title)
for b,v in zip(bars,vals):ax.text(v+.7,b.get_y()+b.get_height()/2,f'{v:.2f}%',va='center',fontsize=10)
ax.axvline(vals[0],color='#637185',ls='--',lw=1)
fig.suptitle('固定8%入场、85%退出、50只:双侧平滑未改善本样本表现\n2024-01-02至2026-10-23,单边30基点费用',fontsize=14)
fig.savefig(FIG/'均值中位数对照.png');plt.close(fig)
chosen=[('baseline','#637185','-'),('mean_3_both','#3d7eff','-'),('median_3_both','#d8a632','--'),('mean_10_both','#9b74b1','-.')]
fig,axes=plt.subplots(2,1,figsize=(12.5,8.3),layout='constrained')
for k,c,style in chosen:
rows=accounts[k]['daily'];dates=[dt.date.fromisoformat(r['date']) for r in rows]
axes[0].plot(dates,[r['nav'] for r in rows],label=name(k),color=c,ls=style,lw=1.7)
axes[0].set_title('短窗口降低收益,长窗口漏掉退出后继续承受回撤');axes[0].set_ylabel('累计净值(起点1)');axes[0].legend(ncol=4,fontsize=10);axes[0].grid(alpha=.2)
for k,c,style in chosen:
rows=[r for r in signals[k] if '2025-12-15'<=r['date']<='2026-02-13']
axes[1].plot([dt.date.fromisoformat(r['date']) for r in rows],[r['exit_reading']*100 for r in rows],label=name(k),color=c,ls=style,lw=1.8)
axes[1].axhline(85,color='#ce5454',ls=':',lw=1.7,label='85%退出线');axes[1].set_ylim(30,100);axes[1].yaxis.set_major_formatter(ticker.PercentFormatter());axes[1].set_title('2026年初:10日均值最高84.72%,没有越过退出线');axes[1].set_ylabel('用于退出判断的历史分位');axes[1].legend(ncol=3,fontsize=10);axes[1].grid(alpha=.2)
fig.savefig(FIG/'净值与退出延迟.png');plt.close(fig)
main=table(['两侧都平滑','累计收益','年化收益','最大回撤','60基点累计收益','已结束 / 总轮次'],[[name(k),pct(g[k]['total_return']),pct(g[k]['annualized_return']),pct(g[k]['max_drawdown']),pct(stress[k]['total_return']),f"{g[k]['closed_cycles']} / {g[k]['cycle_count']}"] for k in both])
one=table(['方法','仅平滑入场:收益 / 回撤','仅平滑退出:收益 / 回撤','入退场都平滑:收益 / 回撤'],[[name(f'{m}_{w}_both')]+[pct(g[f'{m}_{w}_{scope}']['total_return'])+' / '+pct(g[f'{m}_{w}_{scope}']['max_drawdown']) for scope in ('entry','exit','both')] for m in p['methods'] for w in p['windows']])
annual=table(['两侧都平滑','2024','2025','2026至10月23日'],[[name(k)]+[pct(g[k]['annual'][y]['return']) for y in ('2024','2025','2026')] for k in both])
cycle_names=['baseline','mean_3_both','median_3_both','mean_5_both','median_5_both','mean_10_both','median_10_both','mean_20_both','median_20_both']
cycles=table(['方法','2024首次退出日','该轮收益','2025年10月入场后的退出日','2026低位新入场日'],[[name(k),g[k]['cycles'][0]['exit_date'],pct(g[k]['cycles'][0]['return_value']),g[k]['cycles'][1]['exit_date'] or '截至末日仍持有',next((c['entry_date'] for c in g[k]['cycles'] if c['reason']=='low_liquidity'),'已有持仓,未新增轮次')] for k in cycle_names])
costs=table(['两侧都平滑','买卖笔数','持仓交易日','日均双边换手','累计费用 / 初始本金','低位信号连续段数'],[[name(k),g[k]['trade_count'],g[k]['exposure_days'],plain(g[k]['mean_turnover']),plain(g[k]['fees_initial_capital']),g[k]['signals']['low_episodes']] for k in both])
first2025=sum(r['date'][:4]=='2024' for r in accounts['baseline']['daily'])-1
base_eval=accounts['baseline']['daily'][-1]['nav']/accounts['baseline']['daily'][first2025]['nav']-1
seq_eval=seq['accounts']['30']['daily'][-1]['nav']/seq['accounts']['30']['daily'][first2025]['nav']-1
summary={'primary_candidates':25,'smoothing_candidates':24,'unique_paths':manifest['unique_primary_paths'],'baseline_return':g['baseline']['total_return'],'baseline_drawdown':g['baseline']['max_drawdown'],'beat_baseline_30':sum(v['total_return']>g['baseline']['total_return'] for k,v in g.items() if k not in ('baseline','percentile_1')),'beat_baseline_60':sum(v['total_return']>stress['baseline']['total_return'] for k,v in stress.items() if k not in ('baseline','percentile_1')),'improved_drawdown_30':sum(v['max_drawdown']>g['baseline']['max_drawdown']+1e-12 for k,v in g.items() if k not in ('baseline','percentile_1')),'sequential_evaluation_return':seq_eval,'baseline_evaluation_return':base_eval,'jan_2026_mean10_peak':max(r['exit_reading'] for r in signals['mean_10_both'] if '2026-01-01'<=r['date']<='2026-02-13')}
(OUT/'interpretation.json').write_text(json.dumps(summary,ensure_ascii=False,indent=2)+'\n')
text=f'''# 微盘轮动历史分位多日平滑实验
2026年10月28日。基线为用户已经选定的 **8%入场、85%退出、50只**。区间沿用2024年1月2日至2026年10月23日,共678个交易日。只进行开发机离线实验,155运行参数、页面和容器未改。
## 结论
**本次固定8%和85%的实验不支持增加多日平滑。** 24种平滑组合在30基点和60基点费用下,累计收益均低于未平滑基线,最大回撤也均未改善。短窗口主要增加入退场延迟;10至20日窗口会把2026年初的高位信号压到85%以下,从而漏掉退出。
基线累计绝对收益为{pct(g['baseline']['total_return'])},最大回撤{pct(g['baseline']['max_drawdown'])}。入退场都取3日中位数时,收益降为{pct(g['median_3_both']['total_return'])},回撤不变;都取3日均值时降为{pct(g['mean_3_both']['total_return'])}。都取10日均值时,收益{pct(g['mean_10_both']['total_return'])},最大回撤扩大到{pct(g['mean_10_both']['max_drawdown'])}。
这是同一历史、同一费用下的比较,足以说明这批固定阈值方案没有改善历史结果。有效独立交易轮次只有两到三轮,不能据此断言平滑在所有市场都无效,也不能把各组合当作独立样本计算胜率或显著性。
## 固定实验口径
实验方案在北京时间{p['frozen_at'][11:19]}冻结,见[protocol.json](protocol.json)及摘要。没有根据结果继续搜索别的窗口。
平滑对象是**每日活跃度历史百分位**。例如3日均值是信号日及此前两个交易日百分位的平均数;3日中位数取这三个读数的中间值。这里不是平均8%或85%常数,也不是先平滑原始活跃度再重算其历史分位。窗口包括当日收盘已知数据,信号次日开盘执行。
- 方法为均值和中位数,窗口为3、5、10、20个交易日。
- 每组分别测只平滑入场、只平滑退出、两侧同时平滑。共24组,加未平滑基线为25组。
- 入场低于8%,退出高于85%。未平滑的一侧保持原有口径。
- 成长主线瓦解不平滑,仍可独立触发入场。高位退出优先,同时否决新增入场。
- 50只资格名单、按市值选股、日频成员调整、保留成员不重新等权和成交约束不变。
- 2024年初空仓开始。使用2023年12月的真实读数预热完整窗口,不用短窗口凑数、不使用未来值。
当前原始活跃度本身已经用了近5日平均成交额。因此本轮是在已有平滑之上,再对历史百分位进行一次平滑。
原系统用原始活跃度与此前756日分布的线性插值阈值比较。为了避免把离散百分位边界差异误当成平滑效果,另设单日百分位8%/85%对照。实测678天的入退场布尔值、逐笔成交与净值均和基线一致。共运行52条固定方法账户路径,另有两条连续顺序选择账户;25组主比较只有{manifest['unique_primary_paths']}条不同净值路径。
## 入场和退出都平滑
{main}
所有收益都是独立微盘组合的绝对收益,已扣每边30基点交易摩擦。60基点列把同一摩擦假设加倍。空仓不计利息,2026年不是完整年度;未结束轮次按末日估值,不当作已实现利润。不是中证2000相对收益,也不是两策略合并账户收益。
![均值与中位数的收益回撤对照](figures/均值中位数对照.png)
## 分开看入场与退出
{one}
仅对入场做3日均值、5日均值或3日中位数,收益均降为73.62%,回撤仍为−21.07%。这三组都把2026年的低位入场从7月27日推迟到7月28日,最后一轮估值收益由17.61%降为13.80%。其他两轮由成长瓦解入场,保持原日期,所以平滑入场没有改变它们。
退出的影响更大。3日均值和3日中位数都推迟2024年的首次退出,长期持仓产生的调入调出也继续发生。仅平滑退出没有换来最大回撤改善。20日中位数仅平滑入场时,2024年10月退出后的滞后低位信号曾被当日高位条件否决。2026年它的最低读数为8.60%,始终没有低于8%,所以没有第三轮入场。这是两个不同现象。
## 逐轮解释:延迟和漏掉退出
{cycles}
**2024年退出延迟。** 基线9月30日收盘触发,10月8日执行。3日中位数延到10月9日退出,该轮收益由33.66%降为24.99%;3日均值延到10月10日退出,降为17.05%。这里的延迟是一个或两个交易日,不能按跨国庆假期的自然日理解。
**2026年高位信号被压低。** 当年1月至2月中旬,单日百分位最高88.49%,基线1月13日触发、1月14日退出。10日均值最高84.72%,10日中位数最高84.92%,均未达到85%。两者一直持有2025年10月进入的组合至回测末,经历2026年后续下跌,最大回撤扩大至31.73%。7月低位时它们仍在持仓,因此没有新的独立入场轮次;不能简单说成低位入场信号失效。
20日均值在2024年确实比基线多赚了一点,该轮34.04%对33.66%。但它也漏掉了2026年初退出,全区间结果仍差。这说明平滑效果取决于具体行情路径,不能只挑有利的一轮解释。
![净值与2026年初的退出信号](figures/净值与退出延迟.png)
## 年份差异
{annual}
2025年各组合收益相同,均为6.47%,不构成多份独立支持。主要差异集中在2024年退出,以及2026年退出和再次入场。基线两个已结束轮次收益为33.66%和14.15%,均值和中位数均为23.90%,样本仅两轮。
## 费用与换手
{costs}
低位信号连续段数有所减少,但多数短窗口的实际宏观轮次仍是三轮。原策略本来在入场后保持持仓直至高位退出,因此低位指标的小幅反复并不会每天产生新一轮买卖。信号更平滑不等于委托更少。
10至20日双侧平滑的在场时间明显增加,日频成员调整继续发生,买卖笔数反而从基线1330笔升到1856至1910笔。60基点费用压力下,24组依旧全部低于基线。费用占初始本金的比例受累计持仓时间及账户资产规模共同影响,不是固定年费率;日均双边换手包含空仓日。
## 按时间顺序选择也没有改善
沿用前次公开的诊断评分“年化收益减回撤绝对值”,只看2024年、只看2024至2025年分别选下一年的方法,两次都选中20日均值双侧平滑。现金和持仓连续,2024先以基线运行,年末收盘后才更换下一日信号,没有年初清零或拼接独立净值。
2025年至2026年10月23日,连续选择账户收益{pct(seq_eval)},固定8%/85%/50只基线为{pct(base_eval)}。选择账户最大回撤31.73%,固定基线在同段为12.78%。2024年微小优势在后面没有延续。
这个顺序检验仍是已经看过的历史回溯,不是未触碰的留出样本。多次筛选同一历史可能造成挑选偏差,方法依据可参见[Bailey等人的回测过拟合研究](https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf)。本轮没有据此报告统计显著性或估计过拟合概率。
## 可信边界和后续判断
目前保留线上未平滑的8%/85%/50只更符合这轮证据。这里检验的是**固定8%和85%后平滑每日百分位**。若重新对平滑序列标定阈值,属于另一个实验,不能用本轮结果替代,也不能将来选出的更低退出线冒称本轮预先约定方案。
历史覆盖和账户模型继承上一轮:退市股票及历史简称覆盖仍不完整,2026年6月17日备用市值缺整日但主来源存在,12条行情修复保留。交易以可分割复权单位计,不是整手、现金分红与逐笔最低佣金的真实账户。每边30和60基点为统一摩擦假设。未编造主观量化历史持仓或卖出筹资延迟,未证明09:30至10:00的实际成交容量。
## 复核与复现
9项断网隔离测试通过,覆盖均值和中位数手算、缺失值、完整窗口、阈值边界、成长条件和退出优先、次日成交、残余退出、换方法时账户连续及选参不读未来。
独立复算27条主费用账户的现金、逐笔成交和持仓估值,最大误差{max(x['max_cash_nav_error'] for x in audit['independent_accounts'].values()):.2e}。26种口径各自于2024年末和2025年末截断重跑,共52次,全部与完整回测前缀一致。17,628条平滑读数另用NumPy复核,成长信号未改变。详见[审核记录](results/audit.json)。
本轮仅在开发机研究环境运行。155与188均未修改、未重启。线上成交效果未判收,离线结果不当作线上验收。
开发机仓库根目录可运行以下命令,依赖Python、NumPy与Matplotlib,以及前次冻结本地缓存。预期研究测试为9项OK,审核脚本输出27账户及52次截断通过。
```bash
python research/micro_smoothing_20261028/experiment.py
python scripts/unit_test_guard.py research/micro_smoothing_20261028/test_smoothing.py
python scripts/unit_test_guard.py research/micro_smoothing_20261028/audit_smoothing.py
python research/micro_smoothing_20261028/report.py
```
完整文件:[全部方法及费用对照](results/comparison.csv)、[逐笔与逐日账户](results/accounts_30.json.gz)、[信号读数](results/signals.json.gz)、[逐轮比较](results/cycles_comparison.json)、[连续顺序检验](results/sequential.json.gz)、[源数据摘要](results/coverage.json)、[运行摘要](results/manifest.json)。没有部署或自动调参。
'''
(ROOT/'微盘轮动多日平滑实验.md').write_text(text)
print('报告与2张图完成')

View File

@ -0,0 +1,459 @@
{
"independent_accounts": {
"baseline": {
"days": 678,
"trades": 1330,
"max_cash_nav_error": 2.4424906541753444e-15
},
"mean_3_entry": {
"days": 678,
"trades": 1326,
"max_cash_nav_error": 2.6645352591003757e-15
},
"mean_3_exit": {
"days": 678,
"trades": 1346,
"max_cash_nav_error": 2.886579864025407e-15
},
"mean_3_both": {
"days": 678,
"trades": 1342,
"max_cash_nav_error": 2.886579864025407e-15
},
"mean_5_entry": {
"days": 678,
"trades": 1326,
"max_cash_nav_error": 2.6645352591003757e-15
},
"mean_5_exit": {
"days": 678,
"trades": 1356,
"max_cash_nav_error": 2.886579864025407e-15
},
"mean_5_both": {
"days": 678,
"trades": 1352,
"max_cash_nav_error": 2.6645352591003757e-15
},
"mean_10_entry": {
"days": 678,
"trades": 1316,
"max_cash_nav_error": 2.6645352591003757e-15
},
"mean_10_exit": {
"days": 678,
"trades": 1868,
"max_cash_nav_error": 3.552713678800501e-15
},
"mean_10_both": {
"days": 678,
"trades": 1868,
"max_cash_nav_error": 3.552713678800501e-15
},
"mean_20_entry": {
"days": 678,
"trades": 1276,
"max_cash_nav_error": 1.9984014443252818e-15
},
"mean_20_exit": {
"days": 678,
"trades": 1910,
"max_cash_nav_error": 3.1086244689504383e-15
},
"mean_20_both": {
"days": 678,
"trades": 1910,
"max_cash_nav_error": 3.1086244689504383e-15
},
"median_3_entry": {
"days": 678,
"trades": 1326,
"max_cash_nav_error": 2.6645352591003757e-15
},
"median_3_exit": {
"days": 678,
"trades": 1338,
"max_cash_nav_error": 2.6645352591003757e-15
},
"median_3_both": {
"days": 678,
"trades": 1334,
"max_cash_nav_error": 2.4424906541753444e-15
},
"median_5_entry": {
"days": 678,
"trades": 1322,
"max_cash_nav_error": 2.886579864025407e-15
},
"median_5_exit": {
"days": 678,
"trades": 1346,
"max_cash_nav_error": 2.6645352591003757e-15
},
"median_5_both": {
"days": 678,
"trades": 1338,
"max_cash_nav_error": 2.220446049250313e-15
},
"median_10_entry": {
"days": 678,
"trades": 1310,
"max_cash_nav_error": 2.6645352591003757e-15
},
"median_10_exit": {
"days": 678,
"trades": 1856,
"max_cash_nav_error": 3.1086244689504383e-15
},
"median_10_both": {
"days": 678,
"trades": 1856,
"max_cash_nav_error": 3.1086244689504383e-15
},
"median_20_entry": {
"days": 678,
"trades": 988,
"max_cash_nav_error": 1.7763568394002505e-15
},
"median_20_exit": {
"days": 678,
"trades": 1876,
"max_cash_nav_error": 4.218847493575595e-15
},
"median_20_both": {
"days": 678,
"trades": 1876,
"max_cash_nav_error": 4.218847493575595e-15
},
"percentile_1": {
"days": 678,
"trades": 1330,
"max_cash_nav_error": 2.4424906541753444e-15
},
"sequential": {
"days": 678,
"trades": 1824,
"max_cash_nav_error": 3.1086244689504383e-15
}
},
"prefix_checks": [
{
"method_id": "baseline",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "baseline",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_3_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_3_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_3_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_3_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_3_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_3_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_5_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_5_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_5_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_5_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_5_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_5_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_10_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_10_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_10_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_10_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_10_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_10_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_20_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_20_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_20_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_20_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "mean_20_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "mean_20_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_3_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_3_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_3_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_3_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_3_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_3_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_5_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_5_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_5_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_5_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_5_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_5_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_10_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_10_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_10_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_10_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_10_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_10_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_20_entry",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_20_entry",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_20_exit",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_20_exit",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "median_20_both",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "median_20_both",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
},
{
"method_id": "percentile_1",
"through": "2024-12-31",
"days": 242,
"max_difference": 0.0
},
{
"method_id": "percentile_1",
"through": "2025-12-31",
"days": 485,
"max_difference": 0.0
}
],
"signal_rows_checked": 17628,
"growth_unchanged": true,
"percentile_bridge_exact_account_match": true,
"higher_cost_never_better": true,
"sequential_selection_uses_training_only": true,
"note": "独立余额、逐笔先卖后买、交易约束与估值复算;所有26条账户在2024和2025末截断重跑。未验证历史数据库缺失的股票。"
}

View File

@ -0,0 +1,53 @@
method_id,method,window,scope,cost_bps,total_return,annualized,drawdown,2024,2025,2026,cycles,closed_cycles,trades,fees_initial_capital,mean_turnover,exposure_days,low_signal_days,high_signal_days,entry_vetoes
baseline,raw,1,none,30,0.7944394273468172,0.23164447005605715,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.26096326736961495,3,2,1330,0.11338488145635922,0.04359183626783783,235,87,134,1
mean_3_entry,mean,3,entry,30,0.736214556432127,0.21725235575324398,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.2200483039822374,3,2,1326,0.11176471670196053,0.04344519794768741,234,86,134,1
mean_3_exit,mean,3,exit,30,0.5677346197485786,0.1737718932002148,-0.21073920900975784,0.17053791302865084,0.06469902015438667,0.2579408185311838,3,2,1346,0.10551208156432161,0.0441934966004367,238,87,132,1
mean_3_both,mean,3,both,30,0.516865727507219,0.16005603634129661,-0.21073920900975784,0.17053791302865084,0.06469902015438667,0.21712392571157335,3,2,1342,0.10409660420885423,0.04404685828028628,237,86,132,1
mean_5_entry,mean,5,entry,30,0.736214556432127,0.21725235575324398,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.2200483039822374,3,2,1326,0.11176471670196053,0.04344519794768741,234,84,134,1
mean_5_exit,mean,5,exit,30,0.5935114856926904,0.18061296303280816,-0.21073920900975784,0.1789572847829506,0.06469902015438644,0.26949289162162593,3,2,1356,0.10731685974037847,0.044599176639381344,240,87,132,1
mean_5_both,mean,5,both,30,0.541806201501116,0.16681716633623545,-0.21073920900975784,0.1789572847829506,0.06469902015438644,0.22830116421343183,3,2,1352,0.1058781089497396,0.044452538319230926,239,84,132,1
mean_10_entry,mean,10,entry,30,0.656222957987113,0.19696406964632462,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.16383773158833947,3,2,1316,0.10902480007849784,0.04306693190191851,232,80,134,1
mean_10_exit,mean,10,exit,30,0.26967301283684453,0.08880402923795883,-0.317337772280755,0.18446157475654923,0.06469902015438667,0.006801973088549795,2,1,1868,0.15219022608790572,0.06327058986110638,369,87,125,1
mean_10_both,mean,10,both,30,0.26967301283684453,0.08880402923795883,-0.317337772280755,0.18446157475654923,0.06469902015438667,0.006801973088549795,2,1,1868,0.15219022608790572,0.06327058986110638,369,80,125,1
mean_20_entry,mean,20,entry,30,0.5050246663596121,0.15682096997213657,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.05758979202249259,3,2,1276,0.10227615246400566,0.041780300918996406,225,75,134,1
mean_20_exit,mean,20,exit,30,0.43687306052026575,0.13787538814192257,-0.317337772280755,0.34044034234174636,0.06469902015438667,0.006801973088549795,2,1,1910,0.17071179973450157,0.06482278395115056,378,87,110,1
mean_20_both,mean,20,both,30,0.43687306052026575,0.13787538814192257,-0.317337772280755,0.34044034234174636,0.06469902015438667,0.006801973088549795,2,1,1910,0.17071179973450157,0.06482278395115056,378,75,110,1
median_3_entry,median,3,entry,30,0.736214556432127,0.21725235575324398,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.2200483039822374,3,2,1326,0.11176471670196053,0.04344519794768741,234,86,134,1
median_3_exit,median,3,exit,30,0.6740587847602566,0.20154151309057,-0.21073920900975784,0.24992409526225123,0.06469902015438644,0.2579408185311831,3,2,1338,0.1092305751892581,0.043893105246112174,237,87,134,1
median_3_both,median,3,both,30,0.6197399511675361,0.18750116036184172,-0.21073920900975784,0.24992409526225123,0.06469902015438644,0.21712392571157313,3,2,1334,0.10771909979559142,0.043746466925961756,236,86,134,1
median_5_entry,median,5,entry,30,0.7138194862932168,0.21163401689257855,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.20431115488432372,3,2,1322,0.11074783694682656,0.0432652107252753,233,87,134,1
median_5_exit,median,5,exit,30,0.5821316284527696,0.1776016463369361,-0.21073920900975784,0.17053791302865084,0.06469902015438667,0.2694928916216248,3,2,1346,0.10586990002462039,0.044193496600436685,239,87,134,1
median_5_both,median,5,both,30,0.5110501772312639,0.15846922365964522,-0.21073920900975784,0.17053791302865084,0.06469902015438667,0.21245756318937703,3,2,1338,0.10354485552292991,0.043866871057874145,237,87,134,1
median_10_entry,median,10,entry,30,0.6643920056573274,0.19906451810556858,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.16957817001415032,3,2,1310,0.10857526522585015,0.042873437049250784,231,83,134,1
median_10_exit,median,10,exit,30,0.26177192124531934,0.08638476581477961,-0.3173377722807549,0.1770907483357509,0.064699020154386,0.006801973088549573,2,1,1856,0.15040222508597476,0.06280745888868378,367,87,128,1
median_10_both,median,10,both,30,0.26177192124531934,0.08638476581477961,-0.3173377722807549,0.1770907483357509,0.064699020154386,0.006801973088549573,2,1,1856,0.15040222508597476,0.06280745888868378,367,83,128,1
median_20_entry,median,20,entry,30,0.5257186442225512,0.16246411499774838,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.07213157345197674,2,2,988,0.07337741017520176,0.03281565495929808,176,78,134,6
median_20_exit,median,20,exit,30,0.3174139816997614,0.10321969982085588,-0.3173377722807549,0.22899850874511185,0.06469902015438644,0.006801973088550017,2,1,1876,0.1571026557213439,0.06359094344228015,372,87,128,1
median_20_both,median,20,both,30,0.3174139816997614,0.10321969982085588,-0.3173377722807549,0.22899850874511185,0.06469902015438644,0.006801973088550017,2,1,1876,0.1571026557213439,0.06359094344228015,372,78,128,1
percentile_1,mean,1,both,30,0.7944394273468172,0.23164447005605715,-0.21073920900975784,0.3365940180545697,0.064699020154386,0.26096326736961495,3,2,1330,0.11338488145635922,0.04359183626783783,235,87,134,1
baseline,raw,1,none,60,0.6411539844573868,0.19307192910044213,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.22841238847383938,3,2,1330,0.21186387454741812,0.04287440642728854,235,87,134,1
mean_3_entry,mean,3,entry,60,0.5883786538058824,0.17925644204519453,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.1889097759267626,3,2,1326,0.20896300650828656,0.042732882740692646,234,86,134,1
mean_3_exit,mean,3,exit,60,0.43169197108905766,0.136411633001581,-0.22079679215520343,0.11608466021385411,0.04684852462421096,0.22537384143739247,3,2,1346,0.1971910207756919,0.04345562606128514,238,87,132,1
mean_3_both,mean,3,both,60,0.3856524051001837,0.12325225859822564,-0.22079679215520343,0.11608466021385411,0.04684852462421096,0.18596894082110804,3,2,1342,0.19466039322937626,0.043314102374689245,237,86,132,1
mean_5_entry,mean,5,entry,60,0.5883786538058824,0.17925644204519453,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.1889097759267626,3,2,1326,0.20896300650828656,0.042732882740692646,234,84,134,1
mean_5_exit,mean,5,exit,60,0.4543546235424656,0.1427893342717712,-0.22079679215520343,0.12337080249839505,0.04684852462421163,0.23669708939781198,3,2,1356,0.20044106433808487,0.04384919136769453,240,87,132,1
mean_5_both,mean,5,both,60,0.40758628439275735,0.12955610761593084,-0.22079679215520343,0.12337080249839505,0.04684852462421163,0.1969280619775715,3,2,1352,0.1978703787808777,0.043707667681098634,239,84,132,1
mean_10_entry,mean,10,entry,60,0.5163692121375727,0.15992071177574618,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.13501033012815133,3,2,1316,0.20406599997783828,0.04236658758431154,232,80,134,1
mean_10_exit,mean,10,exit,60,0.11702613699070996,0.04022421464748449,-0.3292787479659556,0.12656195461680575,0.04684852462421207,-0.05283768708561132,2,1,1868,0.27848982091233504,0.062029279913847886,369,87,125,1
mean_10_both,mean,10,both,60,0.11702613699070996,0.04022421464748449,-0.3292787479659556,0.12656195461680575,0.04684852462421207,-0.05283768708561132,2,1,1868,0.27848982091233504,0.062029279913847886,369,80,125,1
mean_20_entry,mean,20,entry,60,0.3814934210231853,0.12204972667014524,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.03405509114434402,3,2,1276,0.1919830188067268,0.041116539254072805,225,75,134,1
mean_20_exit,mean,20,exit,60,0.2615508615976849,0.08631693873337931,-0.3292787479659557,0.27232045645662684,0.04684852462421252,-0.05283768708561121,2,1,1910,0.31158565221730405,0.06354467326784204,378,87,110,1
mean_20_both,mean,20,both,60,0.2615508615976849,0.08631693873337931,-0.3292787479659557,0.27232045645662684,0.04684852462421252,-0.05283768708561121,2,1,1910,0.31158565221730405,0.06354467326784204,378,75,110,1
median_3_entry,median,3,entry,60,0.5883786538058824,0.17925644204519453,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.1889097759267626,3,2,1326,0.20896300650828656,0.042732882740692646,234,86,134,1
median_3_exit,median,3,exit,60,0.5301578063495205,0.16366822526668257,-0.22079679215520343,0.19284433373898424,0.04684852462421141,0.2253738414373927,3,2,1338,0.20413815863454177,0.043165108062051524,237,87,134,1
median_3_both,median,3,both,60,0.4809518299793163,0.1501932260561183,-0.22079679215520343,0.19284433373898424,0.04684852462421141,0.18596894082110826,3,2,1334,0.2014334850223733,0.04302358437545563,236,86,134,1
median_5_entry,median,5,entry,60,0.5684530271565775,0.17396353237018958,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.17399533958750402,3,2,1322,0.20714849495400012,0.04255809182785624,233,87,134,1
median_5_exit,median,5,exit,60,0.44492173219817954,0.1401425662922715,-0.22079679215520343,0.11608466021385411,0.04684852462421096,0.23669708939781242,3,2,1346,0.1978450759189201,0.043455626061285155,239,87,134,1
median_5_both,median,5,both,60,0.3809136048984836,0.12188189318919429,-0.22079679215520343,0.11608466021385411,0.04684852462421096,0.18191303918568535,3,2,1338,0.19369351274165603,0.043139311461852854,237,87,134,1
median_10_entry,median,10,entry,60,0.5244423268753853,0.16211750020587945,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.14105309896728535,3,2,1310,0.2032727180863475,0.04217793036132974,231,83,134,1
median_10_exit,median,10,exit,60,0.11126083971648804,0.03830786601672953,-0.3292787479659557,0.12074744021010364,0.04684852462421274,-0.052837687085611096,2,1,1856,0.27542866954396944,0.06158078227150237,367,87,128,1
median_10_both,median,10,both,60,0.11126083971648804,0.03830786601672953,-0.3292787479659557,0.12074744021010364,0.04684852462421274,-0.052837687085611096,2,1,1856,0.27542866954396944,0.06158078227150237,367,83,128,1
median_20_entry,median,20,entry,60,0.42573781451669945,0.1347252551000877,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.06717225236301783,2,2,988,0.1393342925435778,0.03230163627321725,176,78,134,6
median_20_exit,median,20,exit,60,0.1589647334958333,0.05397637510760922,-0.3292787479659557,0.16885857211582,0.04684852462421185,-0.05283768708561121,2,1,1876,0.2873422078578523,0.062338283523929366,372,87,128,1
median_20_both,median,20,both,60,0.1589647334958333,0.05397637510760922,-0.3292787479659557,0.16885857211582,0.04684852462421185,-0.05283768708561121,2,1,1876,0.2873422078578523,0.062338283523929366,372,78,128,1
percentile_1,mean,1,both,60,0.6411539844573868,0.19307192910044213,-0.22079679215520343,0.2762075176259775,0.04684852462421141,0.22841238847383938,3,2,1330,0.21186387454741812,0.04287440642728854,235,87,134,1
1 method_id method window scope cost_bps total_return annualized drawdown 2024 2025 2026 cycles closed_cycles trades fees_initial_capital mean_turnover exposure_days low_signal_days high_signal_days entry_vetoes
2 baseline raw 1 none 30 0.7944394273468172 0.23164447005605715 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.26096326736961495 3 2 1330 0.11338488145635922 0.04359183626783783 235 87 134 1
3 mean_3_entry mean 3 entry 30 0.736214556432127 0.21725235575324398 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.2200483039822374 3 2 1326 0.11176471670196053 0.04344519794768741 234 86 134 1
4 mean_3_exit mean 3 exit 30 0.5677346197485786 0.1737718932002148 -0.21073920900975784 0.17053791302865084 0.06469902015438667 0.2579408185311838 3 2 1346 0.10551208156432161 0.0441934966004367 238 87 132 1
5 mean_3_both mean 3 both 30 0.516865727507219 0.16005603634129661 -0.21073920900975784 0.17053791302865084 0.06469902015438667 0.21712392571157335 3 2 1342 0.10409660420885423 0.04404685828028628 237 86 132 1
6 mean_5_entry mean 5 entry 30 0.736214556432127 0.21725235575324398 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.2200483039822374 3 2 1326 0.11176471670196053 0.04344519794768741 234 84 134 1
7 mean_5_exit mean 5 exit 30 0.5935114856926904 0.18061296303280816 -0.21073920900975784 0.1789572847829506 0.06469902015438644 0.26949289162162593 3 2 1356 0.10731685974037847 0.044599176639381344 240 87 132 1
8 mean_5_both mean 5 both 30 0.541806201501116 0.16681716633623545 -0.21073920900975784 0.1789572847829506 0.06469902015438644 0.22830116421343183 3 2 1352 0.1058781089497396 0.044452538319230926 239 84 132 1
9 mean_10_entry mean 10 entry 30 0.656222957987113 0.19696406964632462 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.16383773158833947 3 2 1316 0.10902480007849784 0.04306693190191851 232 80 134 1
10 mean_10_exit mean 10 exit 30 0.26967301283684453 0.08880402923795883 -0.317337772280755 0.18446157475654923 0.06469902015438667 0.006801973088549795 2 1 1868 0.15219022608790572 0.06327058986110638 369 87 125 1
11 mean_10_both mean 10 both 30 0.26967301283684453 0.08880402923795883 -0.317337772280755 0.18446157475654923 0.06469902015438667 0.006801973088549795 2 1 1868 0.15219022608790572 0.06327058986110638 369 80 125 1
12 mean_20_entry mean 20 entry 30 0.5050246663596121 0.15682096997213657 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.05758979202249259 3 2 1276 0.10227615246400566 0.041780300918996406 225 75 134 1
13 mean_20_exit mean 20 exit 30 0.43687306052026575 0.13787538814192257 -0.317337772280755 0.34044034234174636 0.06469902015438667 0.006801973088549795 2 1 1910 0.17071179973450157 0.06482278395115056 378 87 110 1
14 mean_20_both mean 20 both 30 0.43687306052026575 0.13787538814192257 -0.317337772280755 0.34044034234174636 0.06469902015438667 0.006801973088549795 2 1 1910 0.17071179973450157 0.06482278395115056 378 75 110 1
15 median_3_entry median 3 entry 30 0.736214556432127 0.21725235575324398 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.2200483039822374 3 2 1326 0.11176471670196053 0.04344519794768741 234 86 134 1
16 median_3_exit median 3 exit 30 0.6740587847602566 0.20154151309057 -0.21073920900975784 0.24992409526225123 0.06469902015438644 0.2579408185311831 3 2 1338 0.1092305751892581 0.043893105246112174 237 87 134 1
17 median_3_both median 3 both 30 0.6197399511675361 0.18750116036184172 -0.21073920900975784 0.24992409526225123 0.06469902015438644 0.21712392571157313 3 2 1334 0.10771909979559142 0.043746466925961756 236 86 134 1
18 median_5_entry median 5 entry 30 0.7138194862932168 0.21163401689257855 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.20431115488432372 3 2 1322 0.11074783694682656 0.0432652107252753 233 87 134 1
19 median_5_exit median 5 exit 30 0.5821316284527696 0.1776016463369361 -0.21073920900975784 0.17053791302865084 0.06469902015438667 0.2694928916216248 3 2 1346 0.10586990002462039 0.044193496600436685 239 87 134 1
20 median_5_both median 5 both 30 0.5110501772312639 0.15846922365964522 -0.21073920900975784 0.17053791302865084 0.06469902015438667 0.21245756318937703 3 2 1338 0.10354485552292991 0.043866871057874145 237 87 134 1
21 median_10_entry median 10 entry 30 0.6643920056573274 0.19906451810556858 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.16957817001415032 3 2 1310 0.10857526522585015 0.042873437049250784 231 83 134 1
22 median_10_exit median 10 exit 30 0.26177192124531934 0.08638476581477961 -0.3173377722807549 0.1770907483357509 0.064699020154386 0.006801973088549573 2 1 1856 0.15040222508597476 0.06280745888868378 367 87 128 1
23 median_10_both median 10 both 30 0.26177192124531934 0.08638476581477961 -0.3173377722807549 0.1770907483357509 0.064699020154386 0.006801973088549573 2 1 1856 0.15040222508597476 0.06280745888868378 367 83 128 1
24 median_20_entry median 20 entry 30 0.5257186442225512 0.16246411499774838 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.07213157345197674 2 2 988 0.07337741017520176 0.03281565495929808 176 78 134 6
25 median_20_exit median 20 exit 30 0.3174139816997614 0.10321969982085588 -0.3173377722807549 0.22899850874511185 0.06469902015438644 0.006801973088550017 2 1 1876 0.1571026557213439 0.06359094344228015 372 87 128 1
26 median_20_both median 20 both 30 0.3174139816997614 0.10321969982085588 -0.3173377722807549 0.22899850874511185 0.06469902015438644 0.006801973088550017 2 1 1876 0.1571026557213439 0.06359094344228015 372 78 128 1
27 percentile_1 mean 1 both 30 0.7944394273468172 0.23164447005605715 -0.21073920900975784 0.3365940180545697 0.064699020154386 0.26096326736961495 3 2 1330 0.11338488145635922 0.04359183626783783 235 87 134 1
28 baseline raw 1 none 60 0.6411539844573868 0.19307192910044213 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.22841238847383938 3 2 1330 0.21186387454741812 0.04287440642728854 235 87 134 1
29 mean_3_entry mean 3 entry 60 0.5883786538058824 0.17925644204519453 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.1889097759267626 3 2 1326 0.20896300650828656 0.042732882740692646 234 86 134 1
30 mean_3_exit mean 3 exit 60 0.43169197108905766 0.136411633001581 -0.22079679215520343 0.11608466021385411 0.04684852462421096 0.22537384143739247 3 2 1346 0.1971910207756919 0.04345562606128514 238 87 132 1
31 mean_3_both mean 3 both 60 0.3856524051001837 0.12325225859822564 -0.22079679215520343 0.11608466021385411 0.04684852462421096 0.18596894082110804 3 2 1342 0.19466039322937626 0.043314102374689245 237 86 132 1
32 mean_5_entry mean 5 entry 60 0.5883786538058824 0.17925644204519453 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.1889097759267626 3 2 1326 0.20896300650828656 0.042732882740692646 234 84 134 1
33 mean_5_exit mean 5 exit 60 0.4543546235424656 0.1427893342717712 -0.22079679215520343 0.12337080249839505 0.04684852462421163 0.23669708939781198 3 2 1356 0.20044106433808487 0.04384919136769453 240 87 132 1
34 mean_5_both mean 5 both 60 0.40758628439275735 0.12955610761593084 -0.22079679215520343 0.12337080249839505 0.04684852462421163 0.1969280619775715 3 2 1352 0.1978703787808777 0.043707667681098634 239 84 132 1
35 mean_10_entry mean 10 entry 60 0.5163692121375727 0.15992071177574618 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.13501033012815133 3 2 1316 0.20406599997783828 0.04236658758431154 232 80 134 1
36 mean_10_exit mean 10 exit 60 0.11702613699070996 0.04022421464748449 -0.3292787479659556 0.12656195461680575 0.04684852462421207 -0.05283768708561132 2 1 1868 0.27848982091233504 0.062029279913847886 369 87 125 1
37 mean_10_both mean 10 both 60 0.11702613699070996 0.04022421464748449 -0.3292787479659556 0.12656195461680575 0.04684852462421207 -0.05283768708561132 2 1 1868 0.27848982091233504 0.062029279913847886 369 80 125 1
38 mean_20_entry mean 20 entry 60 0.3814934210231853 0.12204972667014524 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.03405509114434402 3 2 1276 0.1919830188067268 0.041116539254072805 225 75 134 1
39 mean_20_exit mean 20 exit 60 0.2615508615976849 0.08631693873337931 -0.3292787479659557 0.27232045645662684 0.04684852462421252 -0.05283768708561121 2 1 1910 0.31158565221730405 0.06354467326784204 378 87 110 1
40 mean_20_both mean 20 both 60 0.2615508615976849 0.08631693873337931 -0.3292787479659557 0.27232045645662684 0.04684852462421252 -0.05283768708561121 2 1 1910 0.31158565221730405 0.06354467326784204 378 75 110 1
41 median_3_entry median 3 entry 60 0.5883786538058824 0.17925644204519453 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.1889097759267626 3 2 1326 0.20896300650828656 0.042732882740692646 234 86 134 1
42 median_3_exit median 3 exit 60 0.5301578063495205 0.16366822526668257 -0.22079679215520343 0.19284433373898424 0.04684852462421141 0.2253738414373927 3 2 1338 0.20413815863454177 0.043165108062051524 237 87 134 1
43 median_3_both median 3 both 60 0.4809518299793163 0.1501932260561183 -0.22079679215520343 0.19284433373898424 0.04684852462421141 0.18596894082110826 3 2 1334 0.2014334850223733 0.04302358437545563 236 86 134 1
44 median_5_entry median 5 entry 60 0.5684530271565775 0.17396353237018958 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.17399533958750402 3 2 1322 0.20714849495400012 0.04255809182785624 233 87 134 1
45 median_5_exit median 5 exit 60 0.44492173219817954 0.1401425662922715 -0.22079679215520343 0.11608466021385411 0.04684852462421096 0.23669708939781242 3 2 1346 0.1978450759189201 0.043455626061285155 239 87 134 1
46 median_5_both median 5 both 60 0.3809136048984836 0.12188189318919429 -0.22079679215520343 0.11608466021385411 0.04684852462421096 0.18191303918568535 3 2 1338 0.19369351274165603 0.043139311461852854 237 87 134 1
47 median_10_entry median 10 entry 60 0.5244423268753853 0.16211750020587945 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.14105309896728535 3 2 1310 0.2032727180863475 0.04217793036132974 231 83 134 1
48 median_10_exit median 10 exit 60 0.11126083971648804 0.03830786601672953 -0.3292787479659557 0.12074744021010364 0.04684852462421274 -0.052837687085611096 2 1 1856 0.27542866954396944 0.06158078227150237 367 87 128 1
49 median_10_both median 10 both 60 0.11126083971648804 0.03830786601672953 -0.3292787479659557 0.12074744021010364 0.04684852462421274 -0.052837687085611096 2 1 1856 0.27542866954396944 0.06158078227150237 367 83 128 1
50 median_20_entry median 20 entry 60 0.42573781451669945 0.1347252551000877 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.06717225236301783 2 2 988 0.1393342925435778 0.03230163627321725 176 78 134 6
51 median_20_exit median 20 exit 60 0.1589647334958333 0.05397637510760922 -0.3292787479659557 0.16885857211582 0.04684852462421185 -0.05283768708561121 2 1 1876 0.2873422078578523 0.062338283523929366 372 87 128 1
52 median_20_both median 20 both 60 0.1589647334958333 0.05397637510760922 -0.3292787479659557 0.16885857211582 0.04684852462421185 -0.05283768708561121 2 1 1876 0.2873422078578523 0.062338283523929366 372 78 128 1
53 percentile_1 mean 1 both 60 0.6411539844573868 0.19307192910044213 -0.22079679215520343 0.2762075176259775 0.04684852462421141 0.22841238847383938 3 2 1330 0.21186387454741812 0.04287440642728854 235 87 134 1

View File

@ -0,0 +1,11 @@
{
"start": "2024-01-02",
"end": "2026-10-23",
"days": 678,
"warmup_start": "2023-12-05",
"warmup_observations": 19,
"cache_sha256": "7ddffade20bdcff6aa211d8d89876c95f04ba69acae5fd7d9c7262a2df98277d",
"observations_sha256": "a07b5d8bd72702083429cfb9c3bb7f2451db9b29d55bbfc2703510eddd3b7dc5",
"engine_sha256": "1f712dbbe8221c086936b8de6beed2f3af19ce780644c9d8b86358125b56abf5",
"calibrate_sha256": "e6871ed556eaaa74a23fc4cdab5ac8a24838b04909b3341c25387d2fa203f3e9"
}

File diff suppressed because it is too large Load Diff

View File

@ -0,0 +1,13 @@
{
"primary_candidates": 25,
"smoothing_candidates": 24,
"unique_paths": 19,
"baseline_return": 0.7944394273468172,
"baseline_drawdown": -0.21073920900975784,
"beat_baseline_30": 0,
"beat_baseline_60": 0,
"improved_drawdown_30": 0,
"sequential_evaluation_return": 0.07194107423688134,
"baseline_evaluation_return": 0.34254635521910193,
"jan_2026_mean10_peak": 0.8472222222222222
}

View File

@ -0,0 +1,41 @@
{
"protocol_sha256": "cec72a585f671b78134d24eb993dd707deb67849376af673f8937f3314e85a8b",
"experiment_sha256": "c65b334a559baf93475aa8f7aede56984718239d9d05aa12e239ee62f4de8967",
"start": "2024-01-02",
"end": "2026-10-23",
"days": 678,
"warmup_start": "2023-12-05",
"warmup_observations": 19,
"cache_sha256": "7ddffade20bdcff6aa211d8d89876c95f04ba69acae5fd7d9c7262a2df98277d",
"observations_sha256": "a07b5d8bd72702083429cfb9c3bb7f2451db9b29d55bbfc2703510eddd3b7dc5",
"engine_sha256": "1f712dbbe8221c086936b8de6beed2f3af19ce780644c9d8b86358125b56abf5",
"calibrate_sha256": "e6871ed556eaaa74a23fc4cdab5ac8a24838b04909b3341c25387d2fa203f3e9",
"completed_at": "2026-10-28T13:06:34.730234+08:00",
"accounts": 52,
"unique_primary_paths": 19,
"bridge_differences": [],
"sequential_selection": [
{
"train_end": "2024-12-31",
"apply_year": "2025",
"method_id": "mean_20_both",
"train": {
"total_return": 0.34044034234174636,
"annualized_return": 0.34178973871676055,
"max_drawdown": -0.21073920900975784,
"utility": 0.1310505297070027
}
},
{
"train_end": "2025-12-31",
"apply_year": "2026",
"method_id": "mean_20_both",
"train": {
"total_return": 0.4271655190666681,
"annualized_return": 0.19507757297313,
"max_drawdown": -0.21073920900975784,
"utility": -0.015661636036627846
}
}
]
}

Binary file not shown.

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@ -0,0 +1,87 @@
"""平滑实验的时间与资金口径测试。经PMS断网隔离入口运行。"""
import copy
import datetime as dt
import sys
import unittest
from pathlib import Path
sys.path.insert(0,str(Path(__file__).resolve().parent))
from experiment import smooth,signal_series,run,load_data,candidates,read,ROOT,choose
def fixture(values):
data=[];obs=[]
for i,v in enumerate(values):
day=str(dt.date(2024,1,2)+dt.timedelta(days=i))
obs.append({'date':day,'activity_percentile':v})
data.append({'date':day,'members':['A'], 'risk':[],
'rows':{'A':{'open':10.,'close':10.,'adj_factor':1.,'vol':100.,'amount':1000.}},
'limits':{'A':{'up_limit':11.,'down_limit':9.}},
'signal':{'date':day,'activity':v,'activity_percentile':v,'growth_break':False,'thresholds':{'8':.08,'85':.85}}})
return data,obs
def spec(method='mean',window=3,scope='both'):
return {'id':'case','method':method,'window':window,'scope':scope}
class SmoothingTests(unittest.TestCase):
def test_manual_mean_and_median(self):
self.assertEqual(smooth([.1,.9,.2,.3],3,'median'),[None,None,.2,.3])
self.assertAlmostEqual(smooth([.1,.9,.2,.3],3,'mean')[2],.4)
self.assertAlmostEqual(smooth([.1,.9,.2,.3],4,'median')[3],.25)
def test_missing_not_filled_or_shortened(self):
self.assertEqual(smooth([.1,None,.2,.3,.4],3,'mean')[:4],[None]*4)
d,o=fixture([.1,None,.2])
with self.assertRaises(ValueError):signal_series(d[-1:],o,spec())
def test_future_does_not_affect_past(self):
v=[.1,.2,.3,.9,.1]
for m in ('mean','median'):
self.assertEqual(smooth(v,3,m)[:3],smooth(v[:3],3,m))
self.assertEqual(smooth(v,3,m)[:3],smooth(v[:3]+[.01,.99],3,m)[:3])
def test_strict_thresholds(self):
d,o=fixture([.08]*3+[.85]*3)
low=signal_series(d[2:3],o,spec())[0]
high=signal_series(d[-1:],o,spec())[0]
self.assertFalse(low['low']);self.assertFalse(high['high'])
def test_growth_unchanged_and_high_veto_scope(self):
d,o=fixture([.1,.1,.9]);d[-1]['signal']['growth_break']=True
entry=signal_series(d[-1:],o,spec(scope='entry'))[0]
both=signal_series(d[-1:],o,spec(scope='both'))[0]
self.assertTrue(entry['high']);self.assertFalse(both['high'])
self.assertTrue(entry['growth_break'] and both['growth_break'])
def test_next_day_only_and_residual_exit(self):
d,o=fixture([.01,.99,.01,.01,.01])
s=signal_series(d,o,spec(window=1));d[2]['rows']['A']['open']=9.;d[3]['rows']['A']['open']=9.
a=run(d,s)
self.assertEqual([(x['date'],x['side']) for x in a.basket.trades],[(d[1]['date'],'buy'),(d[4]['date'],'sell')])
self.assertEqual(a.decisions[2]['state_after_close'],'exiting')
self.assertEqual(a.decisions[3]['state_after_close'],'exiting')
def test_warmup_from_previous_year(self):
p,d,o,_=load_data();ids={r['date'][:10]:i for i,r in enumerate(o)};i=ids[d[0]['date']]
s=signal_series(d[:1],o,spec(window=20))[0]
self.assertAlmostEqual(s['filtered_percentile'],sum(x['activity_percentile'] for x in o[i-19:i+1])/20)
self.assertLess(o[i-1]['date'][:10],'2024-01-01')
def test_switch_does_not_reset_account(self):
d,o=fixture([.01]*6);a=signal_series(d,o,spec(window=1));b=copy.deepcopy(a)
for x in b:x.update(high=True,low=False)
account=run(d,{'baseline':a,'exit':b},switches={d[2]['date']:'exit'})
self.assertEqual([r['positions'] for r in account.basket.records],[0,1,1,0,0,0])
self.assertLess(account.basket.records[-1]['nav'],1.)
self.assertEqual(account.decisions[3]['executed_signal_date'],d[2]['date'])
def test_choice_does_not_read_future(self):
c=[{'id':'baseline','window':1,'method':'raw','scope':'none'},spec()]
nav={'baseline':[1.,1.1,.5],'case':[1.,1.05,100.]};dates=['2024-01-02','2024-12-31','2025-01-02']
first=choose(c,nav,dates,2)
nav['baseline'][-1]=1000.;nav['case'][-1]=.01
self.assertEqual(choose(c,nav,dates,2),first);self.assertEqual(first[0],'baseline')
if __name__=='__main__':unittest.main()

View File

@ -0,0 +1,152 @@
# 微盘轮动历史分位多日平滑实验
2026年10月28日。基线为用户已经选定的 **8%入场、85%退出、50只**。区间沿用2024年1月2日至2026年10月23日,共678个交易日。只进行开发机离线实验,155运行参数、页面和容器未改。
## 结论
**本次固定8%和85%的实验不支持增加多日平滑。** 24种平滑组合在30基点和60基点费用下,累计收益均低于未平滑基线,最大回撤也均未改善。短窗口主要增加入退场延迟;10至20日窗口会把2026年初的高位信号压到85%以下,从而漏掉退出。
基线累计绝对收益为+79.44%,最大回撤-21.07%。入退场都取3日中位数时,收益降为+61.97%,回撤不变;都取3日均值时降为+51.69%。都取10日均值时,收益+26.97%,最大回撤扩大到-31.73%。
这是同一历史、同一费用下的比较,足以说明这批固定阈值方案没有改善历史结果。有效独立交易轮次只有两到三轮,不能据此断言平滑在所有市场都无效,也不能把各组合当作独立样本计算胜率或显著性。
## 固定实验口径
实验方案在北京时间12:57:47冻结,见[protocol.json](protocol.json)及摘要。没有根据结果继续搜索别的窗口。
平滑对象是**每日活跃度历史百分位**。例如3日均值是信号日及此前两个交易日百分位的平均数;3日中位数取这三个读数的中间值。这里不是平均8%或85%常数,也不是先平滑原始活跃度再重算其历史分位。窗口包括当日收盘已知数据,信号次日开盘执行。
- 方法为均值和中位数,窗口为3、5、10、20个交易日。
- 每组分别测只平滑入场、只平滑退出、两侧同时平滑。共24组,加未平滑基线为25组。
- 入场低于8%,退出高于85%。未平滑的一侧保持原有口径。
- 成长主线瓦解不平滑,仍可独立触发入场。高位退出优先,同时否决新增入场。
- 50只资格名单、按市值选股、日频成员调整、保留成员不重新等权和成交约束不变。
- 2024年初空仓开始。使用2023年12月的真实读数预热完整窗口,不用短窗口凑数、不使用未来值。
当前原始活跃度本身已经用了近5日平均成交额。因此本轮是在已有平滑之上,再对历史百分位进行一次平滑。
原系统用原始活跃度与此前756日分布的线性插值阈值比较。为了避免把离散百分位边界差异误当成平滑效果,另设单日百分位8%/85%对照。实测678天的入退场布尔值、逐笔成交与净值均和基线一致。共运行52条固定方法账户路径,另有两条连续顺序选择账户;25组主比较只有19条不同净值路径。
## 入场和退出都平滑
| 两侧都平滑 | 累计收益 | 年化收益 | 最大回撤 | 60基点累计收益 | 已结束 / 总轮次 |
| --- | --- | --- | --- | --- | --- |
| 未平滑基线 | +79.44% | +23.16% | -21.07% | +64.12% | 2 / 3 |
| 3日均值 | +51.69% | +16.01% | -21.07% | +38.57% | 2 / 3 |
| 5日均值 | +54.18% | +16.68% | -21.07% | +40.76% | 2 / 3 |
| 10日均值 | +26.97% | +8.88% | -31.73% | +11.70% | 1 / 2 |
| 20日均值 | +43.69% | +13.79% | -31.73% | +26.16% | 1 / 2 |
| 3日中位数 | +61.97% | +18.75% | -21.07% | +48.10% | 2 / 3 |
| 5日中位数 | +51.11% | +15.85% | -21.07% | +38.09% | 2 / 3 |
| 10日中位数 | +26.18% | +8.64% | -31.73% | +11.13% | 1 / 2 |
| 20日中位数 | +31.74% | +10.32% | -31.73% | +15.90% | 1 / 2 |
所有收益都是独立微盘组合的绝对收益,已扣每边30基点交易摩擦。60基点列把同一摩擦假设加倍。空仓不计利息,2026年不是完整年度;未结束轮次按末日估值,不当作已实现利润。不是中证2000相对收益,也不是两策略合并账户收益。
![均值与中位数的收益回撤对照](figures/均值中位数对照.png)
## 分开看入场与退出
| 方法 | 仅平滑入场:收益 / 回撤 | 仅平滑退出:收益 / 回撤 | 入退场都平滑:收益 / 回撤 |
| --- | --- | --- | --- |
| 3日均值 | +73.62% / -21.07% | +56.77% / -21.07% | +51.69% / -21.07% |
| 5日均值 | +73.62% / -21.07% | +59.35% / -21.07% | +54.18% / -21.07% |
| 10日均值 | +65.62% / -21.07% | +26.97% / -31.73% | +26.97% / -31.73% |
| 20日均值 | +50.50% / -21.07% | +43.69% / -31.73% | +43.69% / -31.73% |
| 3日中位数 | +73.62% / -21.07% | +67.41% / -21.07% | +61.97% / -21.07% |
| 5日中位数 | +71.38% / -21.07% | +58.21% / -21.07% | +51.11% / -21.07% |
| 10日中位数 | +66.44% / -21.07% | +26.18% / -31.73% | +26.18% / -31.73% |
| 20日中位数 | +52.57% / -21.07% | +31.74% / -31.73% | +31.74% / -31.73% |
仅对入场做3日均值、5日均值或3日中位数,收益均降为73.62%,回撤仍为−21.07%。这三组都把2026年的低位入场从7月27日推迟到7月28日,最后一轮估值收益由17.61%降为13.80%。其他两轮由成长瓦解入场,保持原日期,所以平滑入场没有改变它们。
退出的影响更大。3日均值和3日中位数都推迟2024年的首次退出,长期持仓产生的调入调出也继续发生。仅平滑退出没有换来最大回撤改善。20日中位数仅平滑入场时,2024年10月退出后的滞后低位信号曾被当日高位条件否决。2026年它的最低读数为8.60%,始终没有低于8%,所以没有第三轮入场。这是两个不同现象。
## 逐轮解释:延迟和漏掉退出
| 方法 | 2024首次退出日 | 该轮收益 | 2025年10月入场后的退出日 | 2026低位新入场日 |
| --- | --- | --- | --- | --- |
| 未平滑基线 | 2024-10-08 | +33.66% | 2026-01-14 | 2026-07-27 |
| 3日均值 | 2024-10-10 | +17.05% | 2026-01-15 | 2026-07-28 |
| 3日中位数 | 2024-10-09 | +24.99% | 2026-01-15 | 2026-07-28 |
| 5日均值 | 2024-10-11 | +17.90% | 2026-01-16 | 2026-07-28 |
| 5日中位数 | 2024-10-10 | +17.05% | 2026-01-16 | 2026-07-29 |
| 10日均值 | 2024-10-17 | +18.45% | 截至末日仍持有 | 已有持仓,未新增轮次 |
| 10日中位数 | 2024-10-15 | +17.71% | 截至末日仍持有 | 已有持仓,未新增轮次 |
| 20日均值 | 2024-10-30 | +34.04% | 截至末日仍持有 | 已有持仓,未新增轮次 |
| 20日中位数 | 2024-10-22 | +22.90% | 截至末日仍持有 | 已有持仓,未新增轮次 |
**2024年退出延迟。** 基线9月30日收盘触发,10月8日执行。3日中位数延到10月9日退出,该轮收益由33.66%降为24.99%;3日均值延到10月10日退出,降为17.05%。这里的延迟是一个或两个交易日,不能按跨国庆假期的自然日理解。
**2026年高位信号被压低。** 当年1月至2月中旬,单日百分位最高88.49%,基线1月13日触发、1月14日退出。10日均值最高84.72%,10日中位数最高84.92%,均未达到85%。两者一直持有2025年10月进入的组合至回测末,经历2026年后续下跌,最大回撤扩大至31.73%。7月低位时它们仍在持仓,因此没有新的独立入场轮次;不能简单说成低位入场信号失效。
20日均值在2024年确实比基线多赚了一点,该轮34.04%对33.66%。但它也漏掉了2026年初退出,全区间结果仍差。这说明平滑效果取决于具体行情路径,不能只挑有利的一轮解释。
![净值与2026年初的退出信号](figures/净值与退出延迟.png)
## 年份差异
| 两侧都平滑 | 2024 | 2025 | 2026至10月23日 |
| --- | --- | --- | --- |
| 未平滑基线 | +33.66% | +6.47% | +26.10% |
| 3日均值 | +17.05% | +6.47% | +21.71% |
| 5日均值 | +17.90% | +6.47% | +22.83% |
| 10日均值 | +18.45% | +6.47% | +0.68% |
| 20日均值 | +34.04% | +6.47% | +0.68% |
| 3日中位数 | +24.99% | +6.47% | +21.71% |
| 5日中位数 | +17.05% | +6.47% | +21.25% |
| 10日中位数 | +17.71% | +6.47% | +0.68% |
| 20日中位数 | +22.90% | +6.47% | +0.68% |
2025年各组合收益相同,均为6.47%,不构成多份独立支持。主要差异集中在2024年退出,以及2026年退出和再次入场。基线两个已结束轮次收益为33.66%和14.15%,均值和中位数均为23.90%,样本仅两轮。
## 费用与换手
| 两侧都平滑 | 买卖笔数 | 持仓交易日 | 日均双边换手 | 累计费用 / 初始本金 | 低位信号连续段数 |
| --- | --- | --- | --- | --- | --- |
| 未平滑基线 | 1330 | 235 | 4.36% | 11.34% | 6 |
| 3日均值 | 1342 | 237 | 4.40% | 10.41% | 5 |
| 5日均值 | 1352 | 239 | 4.45% | 10.59% | 4 |
| 10日均值 | 1868 | 369 | 6.33% | 15.22% | 4 |
| 20日均值 | 1910 | 378 | 6.48% | 17.07% | 2 |
| 3日中位数 | 1334 | 236 | 4.37% | 10.77% | 5 |
| 5日中位数 | 1338 | 237 | 4.39% | 10.35% | 4 |
| 10日中位数 | 1856 | 367 | 6.28% | 15.04% | 4 |
| 20日中位数 | 1876 | 372 | 6.36% | 15.71% | 1 |
低位信号连续段数有所减少,但多数短窗口的实际宏观轮次仍是三轮。原策略本来在入场后保持持仓直至高位退出,因此低位指标的小幅反复并不会每天产生新一轮买卖。信号更平滑不等于委托更少。
10至20日双侧平滑的在场时间明显增加,日频成员调整继续发生,买卖笔数反而从基线1330笔升到1856至1910笔。60基点费用压力下,24组依旧全部低于基线。费用占初始本金的比例受累计持仓时间及账户资产规模共同影响,不是固定年费率;日均双边换手包含空仓日。
## 按时间顺序选择也没有改善
沿用前次公开的诊断评分“年化收益减回撤绝对值”,只看2024年、只看2024至2025年分别选下一年的方法,两次都选中20日均值双侧平滑。现金和持仓连续,2024先以基线运行,年末收盘后才更换下一日信号,没有年初清零或拼接独立净值。
2025年至2026年10月23日,连续选择账户收益+7.19%,固定8%/85%/50只基线为+34.25%。选择账户最大回撤31.73%,固定基线在同段为12.78%。2024年微小优势在后面没有延续。
这个顺序检验仍是已经看过的历史回溯,不是未触碰的留出样本。多次筛选同一历史可能造成挑选偏差,方法依据可参见[Bailey等人的回测过拟合研究](https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf)。本轮没有据此报告统计显著性或估计过拟合概率。
## 可信边界和后续判断
目前保留线上未平滑的8%/85%/50只更符合这轮证据。这里检验的是**固定8%和85%后平滑每日百分位**。若重新对平滑序列标定阈值,属于另一个实验,不能用本轮结果替代,也不能将来选出的更低退出线冒称本轮预先约定方案。
历史覆盖和账户模型继承上一轮:退市股票及历史简称覆盖仍不完整,2026年6月17日备用市值缺整日但主来源存在,12条行情修复保留。交易以可分割复权单位计,不是整手、现金分红与逐笔最低佣金的真实账户。每边30和60基点为统一摩擦假设。未编造主观量化历史持仓或卖出筹资延迟,未证明09:30至10:00的实际成交容量。
## 复核与复现
9项断网隔离测试通过,覆盖均值和中位数手算、缺失值、完整窗口、阈值边界、成长条件和退出优先、次日成交、残余退出、换方法时账户连续及选参不读未来。
独立复算27条主费用账户的现金、逐笔成交和持仓估值,最大误差4.22e-15。26种口径各自于2024年末和2025年末截断重跑,共52次,全部与完整回测前缀一致。17,628条平滑读数另用NumPy复核,成长信号未改变。详见[审核记录](results/audit.json)。
本轮仅在开发机研究环境运行。155与188均未修改、未重启。线上成交效果未判收,离线结果不当作线上验收。
开发机仓库根目录可运行以下命令,依赖Python、NumPy与Matplotlib,以及前次冻结本地缓存。预期研究测试为9项OK,审核脚本输出27账户及52次截断通过。
```bash
python research/micro_smoothing_20261028/experiment.py
python scripts/unit_test_guard.py research/micro_smoothing_20261028/test_smoothing.py
python scripts/unit_test_guard.py research/micro_smoothing_20261028/audit_smoothing.py
python research/micro_smoothing_20261028/report.py
```
完整文件:[全部方法及费用对照](results/comparison.csv)、[逐笔与逐日账户](results/accounts_30.json.gz)、[信号读数](results/signals.json.gz)、[逐轮比较](results/cycles_comparison.json)、[连续顺序检验](results/sequential.json.gz)、[源数据摘要](results/coverage.json)、[运行摘要](results/manifest.json)。没有部署或自动调参。