import pandas as pd
RW=f'{L}/results'
WS=pd.read_csv(f'{RW}/w20_scan_levels.csv'); WP=pd.read_csv(f'{RW}/w20_scan_special.csv'); WT=pd.read_csv(f'{RW}/w20_touch_levels.csv'); WD_=pd.read_csv(f'{RW}/w20_touch_round_vs_nonround.csv'); WH=pd.read_csv(f'{RW}/w20_holds_summary.csv'); WL=pd.read_csv(f'{RW}/w20_holds_by_level.csv')
nsp=int(WP.special.sum())
WSt=WS[['ix','set','L','n','days','x2','x3','z','r_a','r_b','r_c','r_a_train','r_a_test']].rename(columns={'x2':'P(≥2×) %','x3':'P(≥3×) %','z':'settle≈0 %','r_a':'net/premium (a)','r_b':'net/premium (b)','r_c':'net/premium (c)','r_a_train':'(a) train','r_a_test':'(a) test'})
WPt=WP[['ix','set','metric','at15','at20','at20_ci','at25','smooth_fit_at20','special']]
WTt=WT[['ix','set','L','round','n','hold5','hold15','hold15_pct','up15']].rename(columns={'hold5':'low≥L−2 next 5m %','hold15':'low≥L−2 next 15m %','hold15_pct':'low≥0.9L next 15m %','up15':'close>L after 15m %'})
pr=WH[(WH.L==20)&(WH.set=='raw')].copy()
prt=pr[['ix','N','exit','split','n','mean','median','win','no_top3','random_mean','random_pctile','p','q_bh_primary','counts']]
W20_HTML=f"""<h2 id=why-20>Why 20? Is the ~20 premium level itself special on expiry days?</h2>
<p class=note>Added after Satya's 20:14 IST (27-Oct) follow-up: “but there is some thing with this 20 i think”. Design pre-registered at 20:20 IST in <a href='PREREG.md'>PREREG.md</a> (WHY-20) before any number was computed. Same data and rules as the expiry section above: expiring contract, real 1m bars, spot 15:15–15:29 missing, official-close settlement, buyer only, next-open fill + slippage, full costs.</p>
<div class=verdict><b>Verdict: 20 is not special.</b><ul>
<li><b>Level scan:</b> across 10 levels (5–100), the chance of a later 2× or 3×, of settling at ≈0, and the net expectancy all move smoothly with level. 20 falls on the curve for every metric and index: {nsp} of {len(WP)} pre-registered “is 20 an outlier” checks pass. Cheaper options are simply worse, and every level loses money for a buyer.</li>
<li><b>Round-number support:</b> 20 does not hold better than 17, 18, 22 or 23. The small round-number advantage under a fixed −2 tolerance comes from level 10, where −2 is a wide cushion; with a % tolerance it disappears.</li>
<li><b>“Holds at 20, buy the base breakout”:</b> on Nifty it is positive in 2026 (best: N=20, SL 50% + trail, +9.3 pts per trade, n=51). But it lost −8.5 on train, has p=0.29, and is −3.8 without the best 3 trades. 20 was the <i>worst</i> level on train and the best on test: a flip like that is noise, not structure. 0 of 12 pre-registered rows count.</li></ul></div>
<h3>W1. Level scan: first touch of each level from above (every OTM option, every expiry day)</h3>
<p><img src='charts/w20_scan_NIFTY_raw.png' style='max-width:100%'><br><img src='charts/w20_scan_SENSEX_raw.png' style='max-width:100%'><br><img src='charts/w20_scan_SENSEX_scaled.png' style='max-width:100%'></p>
{tb(WSt,2)}
<p class=note>Band ±20% around each level (20 = 16–24). Net/premium = net P&amp;L ÷ fill; −0.29 means you lose on average 29% of the premium paid. Error bars and CIs are a 95% bootstrap over days. Sensex “scaled” = level × (Sensex open / Nifty open) ≈ ×3.3.</p>
<p><b>Is 20 an outlier vs a smooth curve fitted to the other 9 levels?</b></p>{tb(WPt,3)}
<h3>W2. Round-number support: does the premium bounce at 20 more than at 17/18/22/23?</h3>{tb(WTt,1)}
<p><b>Round {{10,20,25,30,50}} minus non-round {{17,18,22,23}}, and 20 minus mean(17,18,22,23)</b> (percentage points, 95% day-bootstrap CI):</p>{tb(WD_,2)}
<p>The fixed “low ≥ L−2” measure falls steadily with level: 61% at 10 and 33% at 20 (Nifty, 5 min). The round-number edge in that measure therefore comes from level 10, where −2 is a 20% cushion. With a fair tolerance (≥0.9L) every level is at 12–18% (Nifty), with no bump at 20. 20 vs its non-round neighbours: Nifty +1.1 points (CI −1.2..3.5), Sensex −1.5 (−3.0..0.1).</p>
<h3>W3. “Holds at 20” variant (zone 18–22, floor 17, base N minutes, buy the break of the base high)</h3>{tb(prt,2)}
<p class=note>A row counts only if the test mean &gt; 0, p &lt; 0.05, q &lt; 0.10, it is ≥ 90th percentile vs random, the train mean &gt; 0, the mean without the top 3 is &gt; 0, and it beats the same construct at 15 and 25. Random = same window and side, option nearest the level, 30 repeats.</p>
<p><b>Same construct at every level</b> (mean net pts per trade; Nifty and Sensex raw). 20 is the worst Nifty level on train and the best on test:</p>{tb(WL,2)}
<p>Time of day (Nifty, holds-at-20): most signals come before 11:00. That window loses on train (−10 to −15) and is about zero on test (−4.2 to +3.1). The 11:00–13:00 and 13:00–14:30 rows have 3–25 trades each and swing sign between train and test. See <a href='results/w20_holds_windows.csv'>w20_holds_windows.csv</a>. Sensitivity table (all levels, BH within the table): 20 of 114 test rows are positive and none is significant; the significant rows are negative.</p>
<p class=note>Files: <a href='results/w20_scan_levels.csv'>w20_scan_levels.csv</a> · <a href='results/w20_scan_special.csv'>w20_scan_special.csv</a> · <a href='results/w20_touch_levels.csv'>w20_touch_levels.csv</a> · <a href='results/w20_touch_round_vs_nonround.csv'>w20_touch_round_vs_nonround.csv</a> · <a href='results/w20_holds_summary.csv'>w20_holds_summary.csv</a> · <a href='results/w20_holds_by_level.csv'>w20_holds_by_level.csv</a> · event files w20_scan / w20_touch / w20_holds(_random).csv · scripts s18_why20.py, s19_why20_stats.py, ec_lib.py.</p>
"""