import pandas as pd,numpy as np
from scipy.stats import fisher_exact
rng=np.random.default_rng(11)
E=pd.read_csv('results/p2_events.csv'); E['date']=pd.to_datetime(E.date)
T=pd.read_csv('results/p2_option_trades.csv'); T['date']=pd.to_datetime(T.date)
FUT0=pd.Timestamp('2024-10-21'); SPLIT=pd.Timestamp('2025-07-01')
def per(d): return np.where(d<SPLIT,'train','test')
E['period']=per(E.date)
for k,j in [('r40_15',('mfe15',40)),('r40_30',('mfe30',40)),('r80_30',('mfe30',80)),('r80_60',('mfe60',80)),('r120_60',('mfe60',120))]: E[k]=(E[j[0]]>=j[1]).astype(float).where(E[j[0]].notna())
E['hold60']=E.hold60.map({True:1.0,False:0.0,'True':1.0,'False':0.0}); E['hold_day']=E.hold_day.map({True:1.0,False:0.0,'True':1.0,'False':0.0})
M=['r40_15','r40_30','r80_30','r80_60','r120_60','hold60','hold_day','to_1514']
def summ(x): return pd.Series({'n':len(x),**{m:x[m].mean()*(100 if m!='to_1514' else 1) for m in M}})
F=E[E.date>=FUT0]
# outcome table per config and period
cfg=F[F.kind=='signal'].groupby(['X','V','win','oi','period']).apply(summ)
ctrl=F[F.kind!='signal'].groupby(['kind','X','period']).apply(summ)
# by year for signals and controls
yr=F.groupby(['kind','X','V','win','oi','year']).apply(summ)
# price-only, all years 2022-2026
po=E[E.kind.isin(['price_bar','at_low_any'])].copy(); po['period2']=np.where(po.date<pd.Timestamp('2025-01-01'),'train 2022-24','test 2025-26')
pot=po.groupby(['kind','X','period2']).apply(summ); poy=po.groupby(['kind','X','year']).apply(summ)
# option pnl: nearest expiry, scale style (primary)
TT=T[(T.expiry_rule=='nearest')]
rnd=TT[TT.kind=='random']
res=[]
for (X,V,win,oi),g in F[F.kind=='signal'].groupby(['X','V','win','oi']):
    for prd in ['train','test','all']:
        gg=g if prd=='all' else g[g.period==prd]
        c=F[(F.kind=='control_quiet')&(F.X==X)]; c=c if prd=='all' else c[c.period==prd]
        a,b=gg.r40_30.dropna(),c.r40_30.dropna()
        p_f=fisher_exact([[a.sum(),len(a)-a.sum()],[b.sum(),len(b)-b.sum()]],alternative='greater')[1] if len(a) and len(b) else np.nan
        row=dict(X=X,V=V,win=win,oi=oi,period=prd,n=len(gg),r40_30=a.mean()*100,ctrl_r40_30=b.mean()*100,ctrl_n=len(b),p_fisher=p_f,hold60=gg.hold60.mean()*100,ctrl_hold60=c.hold60.mean()*100)
        for style in ['scale','sl_only']:
            tr=TT[(TT.kind=='signal')&(TT.X==X)&(TT.V==V)&(TT.win==win)&(TT.oi==oi)&(TT['style']==style)]
            tr=tr if prd=='all' else tr[per(tr.date)==prd]
            rr=rnd[rnd['style']==style]; rr=rr if prd=='all' else rr[per(rr.date)==prd]
            n=len(tr); m=tr.pnl.mean() if n else np.nan
            if n>=3 and len(rr)>n:
                sims=np.array([rr.pnl.sample(n,replace=True,random_state=int(s)).mean() for s in rng.integers(0,1e9,2000)])
                p=(sims>=m).mean(); rm=rr.pnl.mean()
            else: p=np.nan; rm=np.nan
            ct=TT[(TT.kind=='control_quiet')&(TT.X==X)&(TT['style']==style)]; ct=ct if prd=='all' else ct[per(ct.date)==prd]
            row.update({f'{style}_n':n,f'{style}_avg':m,f'{style}_win%':(tr.pnl>0).mean()*100 if n else np.nan,f'{style}_total':tr.pnl.sum(),f'{style}_rand_avg':rm,f'{style}_p_vs_random':p,f'{style}_ctrl_avg':ct.pnl.mean(),f'{style}_ctrl_n':len(ct)})
        res.append(row)
Rz=pd.DataFrame(res)
def holm(p):
    p=np.asarray(p,float); o=np.argsort(np.where(np.isnan(p),2,p)); m=np.isfinite(p).sum(); adj=np.full(len(p),np.nan); run=0
    for r,i in enumerate(o):
        if not np.isfinite(p[i]): continue
        run=max(run,min(1,(m-r)*p[i])); adj[i]=run
    return adj
for prd in ['train','test','all']:
    idx=Rz.period==prd
    for c in ['p_fisher','scale_p_vs_random','sl_only_p_vs_random']: Rz.loc[idx,c+'_holm']=holm(Rz.loc[idx,c].values)
Rz.to_csv('results/p2_config_stats.csv',index=False)
cfg.to_csv('results/p2_outcomes_by_config.csv'); ctrl.to_csv('results/p2_outcomes_controls.csv'); yr.to_csv('results/p2_outcomes_by_year.csv'); pot.to_csv('results/p2_priceonly_train_test.csv'); poy.to_csv('results/p2_priceonly_by_year.csv')
# random benchmark summary
rs=TT[TT.kind.isin(['random','control_quiet','price_bar'])].groupby(['kind','style']).pnl.agg(['count','mean','median',lambda x:(x>0).mean()*100]).round(2)
rs.to_csv('results/p2_benchmarks.csv')
pd.set_option('display.width',250)
print(Rz[Rz.period!='all'][['X','V','win','oi','period','n','r40_30','ctrl_r40_30','p_fisher','p_fisher_holm','hold60','ctrl_hold60','scale_n','scale_avg','scale_win%','scale_rand_avg','scale_p_vs_random','scale_p_vs_random_holm','scale_ctrl_avg']].round(3).to_string())
print(rs); print(ctrl.round(1).to_string()); print(pot.round(1).to_string())