# TEST evaluation of frozen choices + flaw-impact tables + all-exit secondary grid.
import pandas as pd,numpy as np,json
from scipy import stats
T=pd.read_csv('results/trades_candidates.csv',parse_dates=['date']); T=T[T.status=='ok']
RR=pd.read_csv('results/trades_random.csv',parse_dates=['date','src_date'])
CH=json.load(open('results/frozen_choices.json')); cands=['A','B15','B30','C','D','E']
def lotq(d): return 65 if d>=pd.Timestamp('2026-01-01') else 75
def summ(x):
    p=x.pnl; w=p[p>0]; l=p[p<=0]
    rs=(x.pnl*x.date.map(lotq)) if len(x) else pd.Series(dtype=float)
    cum=rs.cumsum().values if len(x) else np.array([0]); dd=(cum-np.maximum.accumulate(np.r_[0,cum])[1:]).min() if len(x) else 0
    mon=(x.assign(rs=rs.values).groupby(x.date.dt.to_period('M')).rs.sum()) if len(x) else pd.Series(dtype=float)
    return dict(n=len(p),mean=p.mean(),median=p.median(),win=(p>0).mean()*100,avg_win=w.mean(),avg_loss=l.mean(),payoff=(w.mean()/abs(l.mean())) if len(w) and len(l) and l.mean()!=0 else np.nan,
                rs_total=rs.sum(),rs_per_trade=rs.mean(),maxdd_rs=dd,worst_month=(str(mon.idxmin()) if len(mon) else ''),worst_month_rs=(mon.min() if len(mon) else np.nan),
                t_p_one=(stats.ttest_1samp(p,0,alternative='greater').pvalue if len(p)>2 else np.nan))
rng=np.random.default_rng(7)
def rand_pct(sub,cand,ex):
    R=RR[(RR.cand==cand)&(RR.ex==ex)]; G={k:g.pnl.values for k,g in R.groupby('src_date')}
    arrs=[G[d] for d in sub.date if d in G]
    if not arrs: return np.nan,np.nan
    means=np.array([np.mean([a[rng.integers(len(a))] for a in arrs]) for _ in range(2000)])
    return (means<sub.pnl.mean()).mean()*100, np.mean(np.concatenate(arrs))
rows=[]
for c in cands:
    ex=CH[c]['ex']
    for sp in ('train','test'):
        x=T[(T.cand==c)&(T.ex==ex)&(T.scen=='REAL')&(T.split==sp)].sort_values('date')
        s=summ(x); pc,rm=rand_pct(x,c,ex); rows.append(dict(cand=c,ex=ex,split=sp,**s,random_mean=rm,random_pctile=pc))
P=pd.DataFrame(rows)
te=P[P.split=='test'].copy(); p=te.t_p_one.values; o=np.argsort(p); m=len(p); q=np.empty(m); q[o]=np.minimum.accumulate((p[o]*m/np.arange(1,m+1))[::-1])[::-1]
P.loc[P.split=='test','q_bh']=np.minimum(q,1); P.round(3).to_csv('results/primary_results.csv',index=False)
pd.set_option('display.width',250); print(P.round(2).to_string())
# per-year / per-month for chosen
Y=T[(T.scen=='REAL')].copy(); Y=Y[[CH[c]['ex']==e for c,e in zip(Y.cand,Y.ex)]]; Y['year']=Y.date.dt.year; Y['ym']=Y.date.dt.to_period('M').astype(str); Y['rs']=Y.pnl*Y.date.map(lotq)
yy=Y.groupby(['cand','year']).agg(n=('pnl','size'),mean=('pnl','mean'),win=('pnl',lambda v:(v>0).mean()*100),rs=('rs','sum')).round(1); yy.to_csv('results/chosen_by_year.csv'); print(yy.to_string())
Y.groupby(['cand','ym']).agg(n=('pnl','size'),mean=('pnl','mean'),rs=('rs','sum')).round(1).to_csv('results/chosen_by_month.csv')
# secondary: all exits x scen x split
G=T.groupby(['cand','ex','scen','split']).pnl.agg(n='size',mean='mean',win=lambda v:(v>0).mean()*100).round(2).reset_index(); G.to_csv('results/grid_all_exits_scen.csv',index=False)
print(G[G.scen=='REAL'].pivot_table(index=['cand','ex'],columns='split',values='mean').round(2).to_string())
# flaw impact (pooled all candidates, train+test)
F=T.groupby(['scen','ex']).agg(n=('pnl','size'),gross=('gross','mean'),cost=('cost','mean'),pnl=('pnl','mean')).round(2); print(F.to_string()); F.to_csv('results/flaw_cost_exit_pooled.csv')