# WHY-20 (PREREG 20:20 IST 27-Oct-2026): level scan, round-number support, holds-at-L variant. Real bars; spot>=15:15 missing.
import sys; sys.path.insert(0,'/workspace/research/popular-strategies/scripts')
from ec_lib import *
from multiprocessing import Pool
LV=[5,10,15,20,25,30,40,50,75,100]; RL=[10,17,18,20,22,23,25,30,50]; HL=[10,15,20,25,30,40,50]
WN=[('W1',0,105),('W2',105,225),('W3',225,315),('W4',315,345)]
def wof(t): return next(w for w,a,b in WN if a<=t<=b)
def otm(typ,K,sp): return (typ=='PE' and K<sp) or (typ=='CE' and K>sp)
def day(fp):
    ix,ds=os.path.basename(fp)[:-8].split('_'); dt=pd.Timestamp(ds); s=SP.get((ix,dt))
    if s is None: return [],[],[],[]
    A,src=load(ix,dt); slip=0.5 if ix=='NIFTY' else 1.5
    sc=DH['SENSEX'].open.get(dt,np.nan)/DH['NIFTY'].open.get(dt,np.nan) if ix=='SENSEX' else 1.0
    sets=[('raw',1.0)]+([('scaled',sc)] if ix=='SENSEX' and not np.isnan(sc) else [])
    E1=[];E2=[];E3=[];E4=[]
    ss=pd.Series(s).ffill().values   # moneyness check only (last known spot); no price filling used in any P&L
    for (typ,K),r in A.items():
        c=r[3]; lo_=r[2]; v=r[4]
        for sn,f in sets:
            # 1. level scan: first close in [0.8L,1.2L] after an earlier close >1.2L
            for L0 in LV:
                L=L0*f; hi=1.2*L; lo=0.8*L; above=False
                for t in range(1,346):
                    if not np.isnan(c[t-1]) and c[t-1]>hi: above=True
                    if not above or np.isnan(c[t]): continue
                    if lo<=c[t]<=hi and (np.isnan(v[t]) or v[t]>0) and otm(typ,K,ss[t]):
                        o=outcome(ix,dt,A,typ,K,t,slip,0.05*L); o.update(ix=ix,date=dt,set=sn,L=L0,typ=typ,K=K,t=t); E1.append(o); break
                    if c[t]<lo: break   # fell through band without a close inside: no first-touch-in-band event
            # 2. round-number touch: first low<=level after earlier close>level
            if sn=='raw' or ix=='SENSEX':
                for L0 in RL:
                    L=L0*f; above=False
                    for t in range(1,360):
                        if not np.isnan(c[t-1]) and c[t-1]>L: above=True
                        if not above or np.isnan(lo_[t]): continue
                        if lo_[t]<=L and otm(typ,K,ss[min(t,359)]):
                            w5=lo_[t+1:t+6]; w15=lo_[t+1:t+16]; c15=c[min(t+15,374)]
                            if np.all(np.isnan(w15)): break
                            E2.append(dict(ix=ix,date=dt,set=sn,L=L0,typ=typ,K=K,t=t,hold5=np.nanmin(w5)>=L-2*f if np.any(~np.isnan(w5)) else np.nan,hold15=np.nanmin(w15)>=L-2*f,
                                hold15_pct=np.nanmin(w15)>=0.9*L,up15=(c15>L) if not np.isnan(c15) else np.nan)); break
        # 3. holds-at-L (raw levels only, plus scaled for Sensex)
    for sn,f in sets:
        for L0 in HL:
            L=L0*f
            for N in (10,20):
                best={}
                for (typ,K),r in A.items():
                    o_,h,l,c,v=r[0],r[1],r[2],r[3],r[4]; above=False; t0=None
                    for t in range(1,346):
                        if not np.isnan(c[t-1]) and c[t-1]>1.1*L: above=True
                        if above and not np.isnan(c[t]) and 0.9*L<=c[t]<=1.1*L and otm(typ,K,ss[t]): t0=t; break
                    if t0 is None or t0+N>=345: continue
                    bl=l[t0+1:t0+N+1]
                    if np.all(np.isnan(bl)) or np.nanmin(bl)<0.85*L: continue
                    bh_=np.nanmax(h[t0+1:t0+N+1]); te=None
                    for t in range(t0+N+1,346):
                        if not np.isnan(l[t]) and l[t]<0.85*L: break
                        if not np.isnan(c[t]) and c[t]>bh_: te=t; break
                    if te is None: continue
                    if typ not in best or te<best[typ][0]: best[typ]=(te,K)
                for typ,(te,K) in best.items():
                    o=outcome(ix,dt,A,typ,K,te,slip,0.05*L); o.update(ix=ix,date=dt,set=sn,L=L0,N=N,typ=typ,K=K,t=te,win=wof(te)); E3.append(o)
                    if L0 in (15,20,25):
                        w,a,b=next(x for x in WN if x[0]==wof(te)); rg=np.random.default_rng(hash((ds,ix,sn,L0,N,typ))%2**32)
                        for rep in range(30):
                            for _ in range(20):
                                u=int(rg.integers(max(a,1),b+1)); cand=[(abs(r[3,u]-L),Kk) for (ty,Kk),r in A.items() if ty==typ and not np.isnan(r[3,u]) and 0.8*L<=r[3,u]<=1.2*L and otm(ty,Kk,ss[u])]
                                if cand: break
                            else: continue
                            q=outcome(ix,dt,A,typ,min(cand)[1],u,slip,0.05*L); q.update(ix=ix,date=dt,set=sn,L=L0,N=N,typ=typ,rep=rep); E4.append(q)
    return E1,E2,E3,E4
if __name__=='__main__':
    fps=sorted(glob.glob(f'{C}/*.parquet')); out=[[],[],[],[]]
    with Pool(7) as p:
        for i,res in enumerate(p.imap_unordered(day,fps)):
            for k in range(4): out[k]+=res[k]
            if i%20==0: print(i,len(out[0]),len(out[2]),flush=True)
    for k,n in enumerate(['w20_scan','w20_touch','w20_holds','w20_holds_random']):
        D=pd.DataFrame(out[k]); D['split']=np.where(D.date<=pd.Timestamp('2025-12-31'),'train','test'); D.to_csv(f'{L}/results/{n}.csv',index=False); print(n,len(D))