385 lines
29 KiB
Python
385 lines
29 KiB
Python
#!/usr/bin/env python3
|
||
"""从固定结果生成中文研究图与报告,不重新选参。"""
|
||
from pathlib import Path
|
||
import json
|
||
import sys
|
||
import numpy as np
|
||
import pandas as pd
|
||
import matplotlib
|
||
matplotlib.use('Agg')
|
||
import matplotlib.pyplot as plt
|
||
from matplotlib import font_manager
|
||
from matplotlib.colors import TwoSlopeNorm
|
||
from engine import HERE, NAMES, dump, sha, protocol, load_data, no_network
|
||
|
||
OUT = HERE/'results'
|
||
FIG = HERE/'figures'
|
||
STYLES = ['000300.SH', '399006.SZ', '932000.CSI']
|
||
PRIMARY = ['market']+STYLES
|
||
TITLE = {'formation': '同步形成', 'direction': '进入方向', 'efficiency': '推进效率', 'full': '三项合用',
|
||
'without_formation': '移除同步形成', 'without_direction': '移除进入方向', 'without_efficiency': '移除推进效率',
|
||
'static': '网格与当前同步', 'grid': '价格成交网格', 'baseline': '原有信息基线'}
|
||
|
||
|
||
def style():
|
||
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': 10, 'savefig.dpi': 160,
|
||
'axes.spines.top': False, 'axes.spines.right': False})
|
||
|
||
|
||
def p(x, digits=2):
|
||
return '—' if x is None or not np.isfinite(x) else f'{x*100:.{digits}f}%'
|
||
|
||
|
||
def n(x, digits=2):
|
||
return '—' if x is None or not np.isfinite(x) else f'{x:.{digits}f}'
|
||
|
||
|
||
def heat(ax, values, support=None, limits=None, cmap='RdBu_r', percentage=True):
|
||
a = np.asarray(values, float).reshape(5,5)
|
||
if limits is None:
|
||
m = max(.01, np.nanmax(np.abs(a)))
|
||
limits=(-m,m)
|
||
im = ax.imshow(a, origin='lower', cmap=cmap, vmin=limits[0], vmax=limits[1], aspect='equal')
|
||
if support is not None:
|
||
good = np.asarray(support, bool).reshape(5,5)
|
||
# 薄样本格覆盖半透明灰色,同时保留数值供探索。
|
||
rgba = np.zeros((5,5,4))
|
||
rgba[~good] = [.8,.82,.84,.78]
|
||
ax.imshow(rgba, origin='lower', aspect='equal')
|
||
else:
|
||
good = np.ones((5,5),bool)
|
||
for i in range(5):
|
||
for j in range(5):
|
||
v=a[i,j]
|
||
label='—' if not np.isfinite(v) else (f'{v*100:+.1f}' if percentage else f'{v:.2f}')
|
||
if not good[i,j]: label+='*'
|
||
rgba=im.cmap(im.norm(v)) if np.isfinite(v) else (1,1,1,1)
|
||
dark=sum(rgba[k]*w for k,w in enumerate([.2126,.7152,.0722]))<.45
|
||
ax.text(j,i,label,ha='center',va='center',fontsize=9,color='white' if dark and good[i,j] else '#172739')
|
||
ticks=['0–20','20–40','40–60','60–80','80–100']
|
||
ax.set_xticks(range(5),ticks,fontsize=8)
|
||
ax.set_yticks(range(5),ticks,fontsize=8)
|
||
ax.set_xlabel('成交活跃度历史百分位')
|
||
ax.set_ylabel('价格位置历史百分位')
|
||
return im
|
||
|
||
|
||
def region_heatmaps(results):
|
||
fig, axes=plt.subplots(3,3,figsize=(15,13),layout='constrained')
|
||
for i,s in enumerate(STYLES):
|
||
rows=results[s]['grid']
|
||
support=[x['display_supported'] for x in rows]
|
||
for j,(field,title,limit,percent) in enumerate([
|
||
('price_sync','价格同步水平',(0.4,1),False),
|
||
('residual_sync','控制波动与动量后的同步差',(-.1,.1),True),
|
||
('median_return','随后20日收益中位数',(-.1,.1),True)]):
|
||
im=heat(axes[i,j],[x.get(field) for x in rows],support,limit,'YlGnBu' if j==0 else 'RdBu_r',percent)
|
||
axes[i,j].set_title(f'{NAMES[s]} · {title}')
|
||
fig.colorbar(im,ax=axes[i,j],shrink=.75,label='相关系数' if j==0 else ('相关系数差,格内数字乘100' if j==1 else '收益比例,格内数字为百分数'))
|
||
fig.suptitle('同一区域的同步程度与后续走势:2024至2026年历史滚动检验',fontsize=17)
|
||
fig.supxlabel('灰格及*:少于30个日期或5次访问;所有网格均为探索展示,相关性增强与上涨方向分别检验。',fontsize=10)
|
||
fig.savefig(FIG/'区域同步与收益热力图.png')
|
||
plt.close(fig)
|
||
|
||
|
||
def support_heatmaps(results):
|
||
fig,axes=plt.subplots(1,3,figsize=(15,5.6),layout='constrained')
|
||
for ax,s in zip(axes,STYLES):
|
||
rows=results[s]['grid']
|
||
a=np.array([r['n'] for r in rows]).reshape(5,5)
|
||
im=ax.imshow(a,origin='lower',vmin=0,vmax=100,cmap='Blues')
|
||
for i in range(5):
|
||
for j in range(5):
|
||
r=rows[i*5+j]
|
||
ax.text(j,i,f"{r['n']}日\n{r['visits']}次",ha='center',va='center',fontsize=10,color='white' if r['n']>65 else '#172739')
|
||
ax.set_xticks(range(5),['低','偏低','中','偏高','高'])
|
||
ax.set_yticks(range(5),['低','偏低','中','偏高','高'])
|
||
ax.set_xlabel('成交活跃度层级');ax.set_ylabel('价格层级');ax.set_title(NAMES[s])
|
||
fig.colorbar(im,ax=axes,label='检验日期数',shrink=.75)
|
||
fig.suptitle('网格覆盖:日期数与访问次数同时展示',fontsize=17)
|
||
fig.supxlabel('一次连续停留算同一次访问;相邻访问和20日未来收益仍可能相关。',fontsize=10)
|
||
fig.savefig(FIG/'网格样本覆盖.png');plt.close(fig)
|
||
|
||
|
||
def incremental(comps):
|
||
pairs=[('formation','static'),('direction','static'),('efficiency','static'),('full','static'),
|
||
('full','without_formation'),('full','without_direction'),('full','without_efficiency')]
|
||
cols=['加入同步形成','加入进入方向','加入推进效率','三项合用','形成的独有贡献','方向的独有贡献','效率的独有贡献']
|
||
a=np.full((4,7),np.nan)
|
||
lookup={(c['target'],c['candidate'],c['reference']):c for c in comps}
|
||
for i,s in enumerate(PRIMARY):
|
||
for j,(cand,ref) in enumerate(pairs):
|
||
a[i,j]=lookup[(s,cand,ref)].get('relative_improvement_pct',np.nan)
|
||
m=max(1,np.nanmax(np.abs(a)))
|
||
fig,ax=plt.subplots(figsize=(15,5.6),layout='constrained')
|
||
im=ax.imshow(a,cmap='RdBu',vmin=-m,vmax=m,aspect='auto')
|
||
for i,s in enumerate(PRIMARY):
|
||
for j,(cand,ref) in enumerate(pairs):
|
||
c=lookup[(s,cand,ref)]
|
||
label=f'{a[i,j]:+.2f}%\n校正p={c["p_holm"]:.2f}'
|
||
ax.text(j,i,label,ha='center',va='center',fontsize=10,color='white' if abs(a[i,j])>m*.65 else '#172739')
|
||
ax.set_xticks(range(7),cols,fontsize=10)
|
||
ax.set_yticks(range(4),[NAMES[s] for s in PRIMARY])
|
||
ax.set_title('概率预测的增量价值:正值表示三分类Brier误差下降',fontsize=17,pad=18)
|
||
fig.colorbar(im,ax=ax,label='相对误差改善(%)',shrink=.85)
|
||
fig.supxlabel('前四列相对“网格与当前同步”;后三列比较完整模型与逐项移除。28项统一Holm校正。',fontsize=10)
|
||
fig.savefig(FIG/'三因素增量比较.png');plt.close(fig)
|
||
fig,axes=plt.subplots(2,2,figsize=(14,9),layout='constrained',sharey=True)
|
||
colors=['#2b6f99','#d18b22','#8a6bb1','#2c8b68']
|
||
for ax,s in zip(axes.flat,PRIMARY):
|
||
for color,(cand,ref) in zip(colors,pairs[:4]):
|
||
c=lookup[(s,cand,ref)]
|
||
years=list(c.get('yearly_improvement_pct',{}))
|
||
ax.plot(years,[c['yearly_improvement_pct'][y] for y in years],marker='o',label=TITLE[cand],color=color)
|
||
ax.axhline(0,color='#8d949e',linewidth=1)
|
||
ax.set_title(NAMES[s]);ax.set_ylabel('Brier相对改善(%)');ax.legend(fontsize=9)
|
||
fig.suptitle('同一因素能否跨年保留优势',fontsize=17)
|
||
fig.supxlabel('2026年仅含截至9月16日、未来20日标签已完整兑现的信号;2026年并非全年。',fontsize=10)
|
||
fig.savefig(FIG/'三因素逐年表现.png');plt.close(fig)
|
||
|
||
|
||
def one_dimension(results):
|
||
fig,axes=plt.subplots(2,3,figsize=(15,8),layout='constrained',sharey=True)
|
||
all_values=[r['median_return']*100 for s in STYLES for dim in ['price_rank','volume_rank']
|
||
for r in results[s]['univariate'][dim] if r['median_return'] is not None]
|
||
low,high=min(0,min(all_values)),max(0,max(all_values));span=max(1,high-low)
|
||
for j,s in enumerate(STYLES):
|
||
for i,dim in enumerate(['price_rank','volume_rank']):
|
||
rows=results[s]['univariate'][dim]
|
||
ax=axes[i,j]
|
||
ax.bar(range(10),[r['median_return']*100 if r['median_return'] is not None else np.nan for r in rows],color='#366b91',alpha=.85)
|
||
ax.axhline(0,color='#777',linewidth=.8)
|
||
ax.set_xticks(range(10),[f'{k*10}–{k*10+10}' for k in range(10)],rotation=45,fontsize=8)
|
||
ax.set_ylabel('未来20日收益中位数(%)')
|
||
ax.set_xlabel(('价格' if i==0 else '成交活跃度')+'历史百分位')
|
||
ax.set_title(NAMES[s])
|
||
ax.set_ylim(low-span*.06,high+span*.18)
|
||
for k,r in enumerate(rows):
|
||
ax.text(k,.97,str(r['n']),transform=ax.get_xaxis_transform(),ha='center',va='top',fontsize=8)
|
||
fig.suptitle('单维分层结果:价格位置与成交活跃度',fontsize=17)
|
||
fig.supxlabel('柱顶行数字为日期数;全部使用同一历史滚动检验区间,尚未加入同步筛选。',fontsize=10)
|
||
fig.savefig(FIG/'单维价格与成交分层.png');plt.close(fig)
|
||
|
||
|
||
def current_chart(results,data):
|
||
fig,axes=plt.subplots(3,1,figsize=(14,11),layout='constrained')
|
||
for ax,s in zip(axes,STYLES):
|
||
c=results[s]['current'];d=data[s].iloc[-120:]
|
||
ax.plot(d.index,d.close,color='#1c4264',lw=1.7,label='指数收盘')
|
||
ax.scatter(d.index[-1],d.close.iloc[-1],color='#172739',zorder=5)
|
||
edges=c['regions']['edges'];b=int(c['price_bin'])
|
||
ax.axhspan(edges[b],edges[b+1],color='#929aa3',alpha=.10)
|
||
ax.axhline(edges[b],color='#717983',ls=':',lw=1)
|
||
label=[f'当前层 {edges[b]:.0f}–{edges[b+1]:.0f}']
|
||
for side,color,word in [('lower','#387e70','下方'),('upper','#b57c33','上方')]:
|
||
zone=c['regions'][side]
|
||
if zone:
|
||
ax.axhspan(zone['low'],zone['high'],color=color,alpha=.14)
|
||
ax.axhline(zone['high'] if side=='lower' else zone['low'],color=color,ls='--',lw=1)
|
||
label.append(f"{word} {zone['low']:.0f}–{zone['high']:.0f}({'同步增强候选' if zone['enhanced'] else '几何参考'})")
|
||
else: label.append(word+'无完整相邻格')
|
||
prob=c['models']['full']['probabilities']
|
||
ax.set_title(f"{NAMES[s]} {c['close']:.2f}点 | 模型原始分数:上涨{prob[2]:.0%} 震荡{prob[1]:.0%} 下跌{prob[0]:.0%}")
|
||
ax.set_ylabel('指数点位')
|
||
ax.text(.01,.03,';'.join(label),transform=ax.transAxes,fontsize=10,bbox={'facecolor':'white','alpha':.85,'edgecolor':'none'})
|
||
ax.margins(x=.04)
|
||
fig.suptitle('截至2026年10月23日的区间快照',fontsize=18)
|
||
fig.supxlabel('灰带为当前价格层,层下沿仅为观察边界;概率校准表现不佳,模型原始分数不作为可信胜率。所有边界按10月23日冻结。',fontsize=10)
|
||
fig.savefig(FIG/'当前区间快照.png');plt.close(fig)
|
||
|
||
|
||
def paths_chart(results,data):
|
||
fig,axes=plt.subplots(1,3,figsize=(15,5.4),layout='constrained',sharey=True)
|
||
stats={}
|
||
for ax,s in zip(axes,STYLES):
|
||
d=data[s];events=results[s]['supplementary']['events'];stats[s]={}
|
||
for direction,color,label in [(1,'#b77434','偏多事件'),(-1,'#386e9a','偏空事件')]:
|
||
curves=[]
|
||
for event in events:
|
||
if event['direction'] != direction:continue
|
||
date=pd.Timestamp(event['date']);t=d.index.get_loc(date);entry=d.open.iloc[t+1]
|
||
curves.append(np.r_[0,d.close.iloc[t+1:t+21].to_numpy()/entry-1,d.open.iloc[t+21]/entry-1])
|
||
if curves:
|
||
a=np.asarray(curves)
|
||
ax.plot(range(22),np.median(a,axis=0)*100,color=color,label=f'{label}中位数 n={len(a)}')
|
||
ax.plot(range(22),np.mean(a,axis=0)*100,color=color,ls='--',alpha=.7,label=f'{label}均值')
|
||
stats[s][str(direction)]={'n':len(a),'mean_path':a.mean(axis=0),'median_path':np.median(a,axis=0)}
|
||
ax.axhline(0,color='#999',lw=.8);ax.legend(fontsize=8);ax.set_title(NAMES[s])
|
||
ax.set_xticks([0,5,10,15,21],['入场','5日收','10日收','15日收','20日后开盘'],rotation=20)
|
||
ax.set_ylabel('指数相对入场开盘变化(%)')
|
||
fig.suptitle('固定完整模型的方向事件路径',fontsize=17)
|
||
fig.supxlabel('方向概率至少50%、比反方向高15个百分点;事件间隔至少20交易日。指数价格路径,不是账户收益。',fontsize=10)
|
||
fig.savefig(FIG/'各风格事件路径.png');plt.close(fig)
|
||
dump(OUT/'event_paths.json',stats)
|
||
|
||
|
||
def calibration_chart(results):
|
||
fig,axes=plt.subplots(1,3,figsize=(15,5),layout='constrained')
|
||
for ax,s in zip(axes,STYLES):
|
||
ax.plot([0,1],[0,1],ls='--',color='#929aa3',label='预测与实际一致')
|
||
for name,color in [('static','#5e8eab'),('full','#bd793a')]:
|
||
rows=[r for r in results[s]['metrics'][name]['calibration'] if r['class']==2 and r['n']>0]
|
||
ax.plot([r['predicted'] for r in rows],[r['observed'] for r in rows],marker='o',color=color,label=TITLE[name])
|
||
if name=='full':
|
||
for r in rows:ax.annotate(str(r['n']), (r['predicted'],r['observed']), xytext=(3,5),textcoords='offset points',fontsize=8)
|
||
ax.set_xlim(0,1);ax.set_ylim(0,1);ax.set_title(NAMES[s])
|
||
ax.set_xlabel('模型预测的上涨概率');ax.set_ylabel('实际上涨达标比例');ax.legend(fontsize=8)
|
||
fig.suptitle('概率校准:模型越确信上涨,实际达标比例没有相应提高',fontsize=17)
|
||
fig.supxlabel('五个固定概率组;数字为完整模型的日期数。上涨达标指20日收益超过信号时半个预测期波动尺度。',fontsize=10)
|
||
fig.savefig(FIG/'上涨概率校准.png');plt.close(fig)
|
||
|
||
|
||
def diagnostics(results,data):
|
||
calendar=data['000001.SH'].index
|
||
records={}
|
||
for s in STYLES:
|
||
pred=pd.read_json(OUT/f'predictions_{s}.json.gz',orient='table')
|
||
dates=pred.index[pred.static_high]
|
||
clusters=[]
|
||
end=-1
|
||
for date in dates:
|
||
pos=calendar.get_loc(date)
|
||
if pos>end:
|
||
clusters.append({'first':str(date.date()),'last':str(date.date()),'dates':1})
|
||
else:
|
||
clusters[-1]['last']=str(date.date());clusters[-1]['dates']+=1
|
||
end=pos+21
|
||
c=results[s]['current'];b=int(c['price_bin']);edges=c['regions']['edges']
|
||
records[s]={'static_high_days':len(dates),'overlapping_label_clusters':clusters,
|
||
'current_band':[edges[b],edges[b+1]],'lower_observation_edge':edges[b],
|
||
'lower_edge_distance_pct':(edges[b]/c['close']-1)*100,
|
||
'current_up_score':c['models']['full']['probabilities'][2],
|
||
'calibration_group':next(r for r in results[s]['metrics']['full']['calibration']
|
||
if r['class']==2 and r['bin']==min(4,int(c['models']['full']['probabilities'][2]*5)))}
|
||
output={'scope':'结果后的集中度和展示诊断;未重选参数,未变更主比较、协议或统计家族。',
|
||
'clustering':'按20日标签实际退出位置连通合并;相连窗口计同一段,不声称这些段相互独立。',
|
||
'styles':records}
|
||
dump(OUT/'diagnostics.json',output)
|
||
return output
|
||
|
||
|
||
def write_report(results,comps,pdata,diag):
|
||
lines=['# 宏观价格成交网格与趋势增量实验', '',
|
||
'本轮以2026年10月23日为行情末日,检验同步形成、进入方向和价格推进效率能否改善20日趋势判断。协议于2026年10月29日北京时间11时10分固定并提交,全部计算使用开发机既有缓存。', '',
|
||
'主结果来自2024、2025、2026三个年度的历史滚动检验;每次只用此前五个完整年训练,并删除退出日期跨入检验年的标签。已有历史参与过假设形成,因此这属于历史外推研究。', '',
|
||
'**本轮结论:三项因素在当前网格交互模型中均未取得总体正增量,完整模型误差增加8.48%至13.05%。** 原有信息基线的误差也低于网格模型,因此下一版应优先检验更简约的网格表示,而不是继续增加指标。', '',
|
||
'值得保留的研究线索是同步形成与高同步停滞的状态差异,以及中证2000相对较强的低位修复倾向;两者目前都不能转成确定趋势信号。当前三个风格落在最低价格层,分位网格在尾部过粗,导致没有完整下方相邻格。这是区间设计需要改进的具体位置。', '',
|
||
'## 三项因素的增量价值', '',
|
||
'Brier分数是上涨、震荡、下跌三种概率的平方误差之和,越小越好。下表正值表示相较网格与当前同步模型的误差下降。四个目标各七项比较统一进行28项Holm校正。', '',
|
||
'| 判断对象 | 检验日数 | 同步形成 | 进入方向 | 推进效率 | 三项合用 |',
|
||
'| --- | ---: | ---: | ---: | ---: | ---: |']
|
||
lookup={(c['target'],c['candidate'],c['reference']):c for c in comps}
|
||
for s in PRIMARY:
|
||
vals=[lookup[(s,k,'static')]['relative_improvement_pct'] for k in ['formation','direction','efficiency','full']]
|
||
lines.append(f"| {NAMES[s]} | {results[s]['n']} | "+' | '.join(f'{x:+.2f}%' for x in vals)+' |')
|
||
accepted=[c for c in comps if c['target'] in PRIMARY and c['accepted']]
|
||
lines += ['', '达到预设综合标准的主比较有'+str(len(accepted))+'项。综合标准要求改善为正、区块区间不跨零、校正后p小于0.05,并在至少三个年度中的两个年度改善。', '',
|
||
'', '',
|
||
'', '',
|
||
'| 对象 | 对比 | Brier绝对改善 | 95%区块区间 | 校正p | 收益预测MSE改善 |',
|
||
'| --- | --- | ---: | --- | ---: | ---: |']
|
||
for c in comps:
|
||
if c['target'] not in PRIMARY:continue
|
||
desc=TITLE[c['candidate']]+' 相对 '+TITLE[c['reference']]
|
||
lo,hi=c['ci95']
|
||
lines.append(f"| {NAMES[c['target']]} | {desc} | {n(c['mean'],4)} | {n(lo,4)} 至 {n(hi,4)} | {c['p_holm']:.3f} | {c['mse_improvement_pct']:+.2f}% |")
|
||
lines += ['', '区块区间对已固定的外推预测误差按60交易日抽样2000次,保留时间相关性;不把每一天或每一网格当成独立试验,也不涵盖全部训练重估不确定性。', '',
|
||
'## 各风格概率表现与覆盖', '',
|
||
'| 对象 | 当前同步模型Brier | 完整模型Brier | 完整模型准确率 | 方向判断覆盖率 |',
|
||
'| --- | ---: | ---: | ---: | ---: |']
|
||
for s in PRIMARY:
|
||
m=results[s]['metrics'];lines.append(f"| {NAMES[s]} | {m['static']['brier']:.4f} | {m['full']['brier']:.4f} | {p(m['full']['accuracy'],1)} | {p(m['full']['coverage'],1)} |")
|
||
lines += ['', '| 对象 | 原有信息基线 | 加网格 | 加当前同步 | 再加三项因素 |',
|
||
'| --- | ---: | ---: | ---: | ---: |']
|
||
for s in PRIMARY:
|
||
m=results[s]['metrics']
|
||
lines.append('| '+NAMES[s]+' | '+' | '.join(f"{m[k]['brier']:.4f}" for k in ['baseline','grid','static','full'])+' |')
|
||
lines += ['', '上表各列为同一日期集合的Brier,逐步扩展使用信息。它说明整套网格相较原有信息的变化;主检验的三因素增量仍以当前同步模型为对照。', '',
|
||
'全市场标签要求沪深300、中证500、中证1000中至少两个出现同方向、超过各自半个预测期波动尺度的变化;它与三指数平均收益的正负并不等同。风格标签分别计算,创业板与宽基的重叠不增加全市场票数。', '',
|
||
'中证2000的历史外推日期均在发布后,五年训练包含发布前回溯数据;它也不等同PMS微盘股票篮子。科创50因自身QRS预热后不足五个完整训练年,不生成同等级外推预测,描述结果保留在机器结果中。', '',
|
||
'## 同步形成与同步高位的状态差异', '',
|
||
'| 风格 | 状态 | 日期数 | 20日均值/中位数 | 相较静态模型预测的平均收益差 |',
|
||
'| --- | --- | ---: | --- | ---: |']
|
||
for s in STYLES:
|
||
for key,label in [('forming','同步形成较强'),('static_high','同步较高且不再增强'),('weak_forming','同步形成较弱')]:
|
||
r=results[s]['synchrony_states'][key]
|
||
lines.append(f"| {NAMES[s]} | {label} | {r['n']} | {p(r['mean_return'])}/{p(r['median_return'])} | {p(r['mean_residual_return'])} |")
|
||
lines += ['', '强弱分位阈值只在对应训练期估计。这张表展示状态差异,不控制三组网格构成完全相同;三因素的正式增量结论以同日期模型比较为准。', '',
|
||
'## 价格与成交分层', '', '', '',
|
||
'', '',
|
||
'', '',
|
||
'二维网格按各自此前252日价格位置与成交活跃度分成五层。价格位置以此前60日均价和20日ATR归一;活跃度使用含当日5日平均成交额除以此前60日成交额中位数。同步残差由每个训练期的基线特征岭回归去除动量、波动、QRS、普通涨跌广度和股汇背景。图中灰格仍保留探索读数,不能据其颜色宣称单格显著。', '',
|
||
'## 截至数据末日的区间判断', '',
|
||
'以下是2026年10月23日收盘快照,不代表10月29日实时判断。概率来自事先固定的完整模型,并与当前同步基线并列;未根据本轮结果更换当前模型。价格百分位是相对均价的ATR偏离位置在过去一年的分位,不是指数绝对点位的历史分位。', '',
|
||
'完整模型的概率校准表现不佳,下列数字只作为原始模型分数,不能直接解释成可信胜率。三指数平均价格相关性为0.925,较此前非重叠60日上升0.022,尚未达到训练期强同步形成门槛0.075。三个风格的成交活跃度分别为常态参照的82.4%、79.7%、86.4%;从价格位置、量能和模型分歧判断,目前更像共同下跌后的分化修复,未形成全市场一致的可确认方向。', '',
|
||
'| 对象 | 收盘点位 | 均价偏离百分位 | 成交百分位 | 5日推进效率 | 原始模型下跌/震荡/上涨 | 模型倾向 |',
|
||
'| --- | ---: | ---: | ---: | ---: | --- | --- |']
|
||
for s in PRIMARY:
|
||
c=results[s]['current'];m=c['models']['full']
|
||
lines.append(f"| {NAMES[s]} | {n(c.get('close'))} | {p(c['price_rank'],1)} | {p(c['volume_rank'],1)} | {p(c['efficiency'],1)} | "+'/'.join(p(v,1) for v in m['probabilities'])+f" | {m['direction']} |")
|
||
lines += ['', '', '',
|
||
'| 风格 | 当前所在价格层 | 层下沿相距现价 | 上方同步区域或参考 |',
|
||
'| --- | --- | ---: | --- |']
|
||
for s in STYLES:
|
||
info=diag['styles'][s];z=results[s]['current']['regions']['upper']
|
||
upper='无' if z is None else f"{z['low']:.0f}至{z['high']:.0f}({'同步增强候选' if z['enhanced'] else '几何参考'})"
|
||
lines.append(f"| {NAMES[s]} | {info['current_band'][0]:.0f}至{info['current_band'][1]:.0f} | {info['lower_edge_distance_pct']:+.2f}% | {upper} |")
|
||
lines += ['', '当前层下沿4016、2553、2521点分别是三个风格的历史分位投影边界,可用于观察是否继续跌出已有状态范围;本轮没有证明它们会形成支撑。由于三个风格都在最下层,其下方没有完整相邻格,双侧首次到达统计的有效分母为零。三个风格同格类比原有74、84、52个日期,均因缺少完整下方区域未进入双侧比较。', '',
|
||
'区间由训练期同成交层的同步残差、样本覆盖筛选,再按快照日均价和ATR投影。没有合格同步增强区域时,只显示相邻几何参考;超出历史有限边界时不外推一个目标。首次到达与触区响应的逐项状态保留在完整结果中。', '',
|
||
'| 风格 | 当前同步基线下跌/震荡/上涨 |',
|
||
'| --- | --- |']
|
||
for s in STYLES:
|
||
c=results[s]['current']
|
||
lines.append(f"| {NAMES[s]} | "+'/'.join(p(v,1) for v in c['models']['static']['probabilities'])+' |')
|
||
lines += ['', '响应定义为触区后五个完整收盘内,在先穿透外侧半个ATR之前反向移动一个ATR;ATR与区间在信号日冻结。开盘越界、触区即穿透和观察不完整分别记账。', '',
|
||
'## 概率校准', '', '', '',
|
||
'| 风格 | 当前上涨原始分数 | 所在概率组历史日数 | 该组平均预测 | 实际上涨达标比例 |',
|
||
'| --- | ---: | ---: | ---: | ---: |']
|
||
for s in STYLES:
|
||
info=diag['styles'][s];g=info['calibration_group']
|
||
lines.append(f"| {NAMES[s]} | {p(info['current_up_score'],1)} | {g['n']} | {p(g['predicted'],1)} | {p(g['observed'],1)} |")
|
||
lines += ['', '此表是历史外推概率组的诊断,不是对当前样本另做概率重标定。它解释为什么当前中证2000的50.7%原始上涨分数只能描述模型倾向。', '',
|
||
'## 方向事件路径与费用压力', '',
|
||
'', '',
|
||
'| 对象 | 事件数 | 10基点方向幅度均值/中位数 | 30基点 | 60基点 |',
|
||
'| --- | ---: | --- | --- | --- |']
|
||
for s in PRIMARY:
|
||
z=results[s]['supplementary'];parts=[]
|
||
for bp in [10,30,60]:
|
||
v=z['summary'][f'{bp}bps'];parts.append(p(v['mean'])+'/'+p(v['median']))
|
||
lines.append(f"| {NAMES[s]} | {z['event_count']} | "+' | '.join(parts)+' |')
|
||
lines += ['', '事件按固定完整模型概率至少50%、比反方向高15个百分点触发,并间隔至少20交易日。偏空只检验价格下跌方向的幅度,不假定股票可卖空。指数开盘标签与统一摩擦不构成账户净收益,也不代表PMS现有09:35调度可按开盘成交。', '',
|
||
'## 下一轮值得优先检验的改动', '',
|
||
'第一,先去掉逐格交互,分别比较单维网格、二维平滑网格和全局三因素。当前完整模型有456列输入,每个训练窗约1190个重叠日标签;网格扩展已经降低外推表现。模型复杂度是否掩盖了原始因素的价值,需要另立对照验证。', '',
|
||
'第二,保留同步形成与高位停滞的区别,但按行情段评估。高同步且不再增强只有21个日期,未来收益窗口相连后合为'+str(len(diag['styles']['000300.SH']['overlapping_label_clusters']))+'段;不能用21个日期当21次独立发现。其明显较差的路径更适合进一步检验为趋势衰减条件。', '',
|
||
'第三,在最低和最高价格层内增加事先固定的ATR细格。当前三个风格均处最下层,分位网格将不同深度的下跌压在同一格,无法形成完整的下方候选区域。补足尾部结构后,再检验支撑反应与反转确认。', '',
|
||
'以上是结果后的研究建议,本轮没有重跑新参数或覆盖原结果。', '',
|
||
'## 可复核资料', '',
|
||
'- [固定协议](protocol.json)与[摘要](protocol.sha256)。',
|
||
'- [完整统计](results/results.json)、[逐项比较](results/comparisons.json)、[数据覆盖](results/data_coverage.json)。',
|
||
'- [独立复核](results/audit.json)与[运行摘要](results/manifest.json)。',
|
||
'- 逐日特征和预测保存在results目录的压缩JSON中;源行情沿用既有本地缓存,协议记录其摘要。',
|
||
'- 研究代码为engine.py、regions.py、run.py、render.py、audit.py及test_engine.py。', '',
|
||
'所有结果在开发机计算。线上应用、参数、持仓、数据库和容器均未改动;本研究代码不进入业务镜像,无受影响业务容器。', '']
|
||
(HERE/'宏观网格首轮研究报告.md').write_text('\n'.join(lines))
|
||
|
||
|
||
def main():
|
||
sys.addaudithook(no_network)
|
||
pdata=protocol();results=json.loads((OUT/'results.json').read_text());comps=json.loads((OUT/'comparisons.json').read_text())
|
||
data,_,_=load_data(pdata)
|
||
FIG.mkdir(exist_ok=True);style()
|
||
region_heatmaps(results);support_heatmaps(results);incremental(comps);one_dimension(results)
|
||
current_chart(results,data);paths_chart(results,data);calibration_chart(results)
|
||
diag=diagnostics(results,data);write_report(results,comps,pdata,diag)
|
||
dump(OUT/'render_manifest.json',{'figures':{x.name:sha(x) for x in FIG.glob('*.png')},'render_code':sha(HERE/'render.py')})
|
||
print('已生成8幅研究图与报告。')
|
||
|
||
|
||
if __name__=='__main__':main()
|