Chart patterns ka sach: 9 lakh candles ki audit

7 min read · data: 437-stock NSE universe · descriptive study, not advice

Har trading group mein koi na koi screenshot aata hai: “hammer bana hai support pe”, “bearish engulfing aa gaya, ab girega”. Candlestick patterns India ke retail trading culture ka sabse popular tool hain. Books, YouTube, courses — sab sikhate hain ki pattern dekho, direction samjho.

Toh humne ek seedha sawaal poocha: agar yeh patterns sach mein direction batate, toh lakhon occurrences mein yeh dikhna chahiye. Yeh ek measurement hai, prediction nahi. Aur result kaafi saaf hai.

Humne kya test kiya

Do alag audits, ek hi discipline:

1. Candle reaction audit (Aug 2026): 480 NSE stocks, 905,399 pattern events, 14 patterns — hammer, doji, inverted hammer, bullish/bearish engulfing, marubozu, three white soldiers types, capitulation candles, waghaira. Har event ke baad 1, 3, 5 aur 10 din ka return mapa.
2. TA direction audit (Jul 2026): NIFTY 50 stocks, 2016–2026. Candlesticks ke saath double top/bottom, head-and-shoulders aur horizontal support/resistance bounce/break bhi. Total 44 pattern × horizon cells.

Sabse important baat: humne return ko zero se compare nahi kiya. Market khud upar jaata rehta hai (isko drift kehte hain). Agar ek stock pattern ke baad 5 din mein +0.8% gaya aur poora market usi 5 din mein +0.8% gaya, toh pattern ne kuch nahi kiya. Isliye har number “same-date market drift minus” hai — yaani pattern ka extra contribution.

Result 1: naked patterns economically dead hain

Bina kisi context ke (sirf pattern dekh ke), 5-day excess return ka sabse bada number tha +0.37 percentage point (marubozu bull, n=6,732) — aur woh bhi statistically significant nahi tha. Baaki sab isse chhote.

Ek line mein samjhein: “bullish” patterns ke baad stock thoda kam chala, “bearish” patterns ke baad thoda zyada. Textbook ka ulta. Yeh short-term mean reversion hai — jo 3 din lagatar gira, woh agle hafte thoda wapas aata hai; jo lagatar chadha, woh thoda saans leta hai. Pattern ka naam isme koi role nahi khelta.

Aur magnitude dekhiye: 0.1–0.2 percentage point, 5 din mein. Brokerage aur slippage (cash delivery mein round-trip roughly 0.3%) isse zyada kha jaata hai.

Result 2: double bottom, head-and-shoulders bhi kuch nahi batate

TA audit ne chart patterns ko ek kadi test diya: pattern ki neckline break ko compare kiya usi stock ki random horizontal line ke break se. Agar neckline special hai, toh usko random line se better perform karna chahiye.

• Double bottom (21 din hold, n=572): raw +0.38%, lekin drift-adjusted −1.02%. Yaani market se kam.
• Inverse head-and-shoulders (n=253): drift-adjusted −1.01%.
• Double top / head-and-shoulders (short side): drift-adjusted ~0, random line se koi farak nahi.
• “Measured move” exit wale numbers bade dikhte the (+2.1%), lekin jab wahi exit rule random lines pe lagaya, toh farak gayab (p-value 0.000 se 0.796 tak chala gaya). Edge pattern mein nahi tha — exit rule ke design mein tha.

Support/resistance bounces (n=3,506) aur breaks (n=1,575) — dono random levels se statistically same.

Result 3: 44 mein se 1 survive kiya — ulti direction mein

Jab aap 44 cells test karte ho, toh sirf chance se ~2 cells “significant” dikh jaate hain. Isliye humne multiple-testing correction (Benjamini-Hochberg) lagaya. Sirf ek cell bachi: piercing line, 5 din (n=149). Textbook mein yeh bullish reversal hai. Data mein iske baad 5 din ka return −1.28% tha (drift-adjusted −0.69%). Yaani jo pattern “upar jaane” ka signal maana jaata hai, woh baad mein girne se zyada juda tha. Aur yeh bhi 11 mein se sirf 7 saal consistent tha — toh isko bhi edge nahi, sirf ek “watch” maana gaya.

Toh kya candles bilkul bekaar hain?

Nahi — ek nuance hai. Candle reaction audit mein jab context joda (stock trend mein hai ya range mein, recent high ke paas hai ya low ke paas), toh kuch cells ne 1–2 percentage point ka farak dikhaya:

• Marubozu bear, recent high ke paas, uptrend ke baad (n=340): agle 5 din −1.8 pp vs drift, win rate 42.6%. Sabse strong cell.
• Doji, recent low ke paas, uptrend mein (n=1,099): agle 10 din +1.75 pp.
• Marubozu bull, recent high ke paas, trend mein (n=2,184): 3 din +0.73 pp.

Inka matlab: candle akela kuch nahi bolta; candle kahan bana aur kis trend mein bana, woh “character” thoda batata hai. Lekin yeh bhi in-sample numbers hain (wahi data jisme dhoondha gaya), winner’s curse lagta hai — asli effect shayad isse chhota ho. Hum inko filter ki tarah dekhte hain, standalone signal ki tarah nahi.

Kya dekhein (kya kharidein nahi)

• Agar koi bole “hammer bana hai toh upar jayega” — base rate yaad rakhiye: 108,902 hammers ke baad 5 din ka excess −0.06 pp tha. Coin flip se bhi thoda kam.
• Pattern se zyada maayne rakhta hai context: trend, location, aur pichhla move.
• Humare /stocks/ pages pe har stock ka structure aur trend roz refresh hota hai — wahan context dekhiye, pattern ka naam nahi.

The limitation

Yeh audit daily (EOD) candles pe hai — intraday patterns pe kuch nahi kehta. NIFTY 50 wale audit mein current constituents peeche tak lagaye gaye (survivorship bias — jo companies index se bahar ho gayi, woh missing hain). Context cells in-sample hain aur kisi forward test se nahi guzre. Base rates past describe karte hain; future predict nahi karte. Lekin folklore ki jagah ek measured prior dena — yahi is site ka kaam hai.

Methodology note. Candle audit: 480 stocks, 905,399 events, 784 cells (min n=300, min 100 distinct dates); null = date-matched universe drift; date-clustered t-stats; BH-FDR 5% → 121 significant cells, 21 with |excess| ≥ 0.5 pp. TA audit: NIFTY 50, Jan 2016 – Jun 2026; signal at close of day t, entry next open, exit close t+h (h = 1, 5, 21); 0.08% friction; three nulls (same-context-no-pattern permutation, date-matched drift, random-level bootstrap); pivots = 10-bar lookback fractal with confirmation lag; 44 cells pooled BH-FDR q<0.10. Source: tietiy internal research (candle_reaction_20260804, ta_direction_audit_2026-07).