Base rate kya hai? Zero se compare galat kyun

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

Is site ke har article mein aap ek line baar baar dekhenge: “excess vs drift” ya “base rate”. Yeh jargon lagta hai, lekin yeh poori site ka sabse important idea hai. Aur iska sabse bada proof hum khud hain — ek hafte mein humne apne hi chaar “edges” isi galti ki wajah se pakde aur maare. Yeh article woh kahani hai.

Base rate: “normally kya hota hai?”

Base rate ka simple matlab hai: kisi bhi random din pe, bina kisi signal ke, kya hota hai?

Example: Agar koi bole “yeh pattern aane ke baad 55% baar stock 10 din mein upar gaya”, toh pehla sawaal hona chahiye — kisi bhi random din ke baad stock 10 din mein kitni baar upar jaata hai? Humare ~500-stock data mein (May 2023 – Oct 2026) jawab hai 50.3%. Toh 55% ka matlab “55% win rate” nahi, matlab “normal se ~5 point zyada”. Aur agar base rate 56% hota, toh woh pattern actually normal se kam hota.

Number ka matlab tabhi banta hai jab aapko pata ho ki normal kya hai.

Drift: market khud chalta hai

Indian stocks lambe samay mein upar gaye hain. Isko drift kehte hain — signal ke bina bhi, bas long rehne se milne wala return. Problem yeh hai: jab market upar ja raha ho, toh lagbhag har “long” signal paisa banata dikhta hai.

Isliye hum har result ko zero se nahi, usi date ke market drift se compare karte hain. Matlab: jis din signal aaya, usi din poore universe ke stocks ne agle N din mein average kya kiya? Signal ka return minus yeh drift = signal ka asli contribution (excess return).

Chaar galat edges jo humne khud pakde (July 2026)

Ek hi hafte mein chaar alag-alag “edges” ek hi galti se bane the — zero se compare.

1. Scanner signals: “72% win rate, +6.29% per trade.” Shandaar lagta tha. Phir humne usi dates pe random 18-stock baskets banaye. Unhone +6.79% diya (p=0.47). Yaani scanner ne stock select nahi kiye the; market upar tha, bas.
2. Setup tiers ke profit factor 4.78 se 19.71. Chhote sample, bear-market bounce pe fit, drift control nahi. Honest test mein gayab.
3. Results ke baad drift (PEAD): “+0.87% net, t=2.74”. Us window mein equal-weight universe khud +0.94% chala tha. Honest benchmark ke baad: training mein +0.11%, test mein −0.14%.
4. Index inclusion: “+1.35% per trade”. Isme test data bhi galat tha, aur drift control bhi nahi. Asli events pe: −5.8%.

Kisi ek ne bhi drift-matched test pass nahi kiya. Isliye ab humara rule hai: koi bhi number site pe aane se pehle drift se compare hota hai.

Ek aur example: results se pehle ka “premium”

US research kehti hai ki earnings se pehle ke ~10 din mein stocks extra chalte hain. Humne 7,940 Indian F&O results events (2014–2026) pe test kiya:

Lagbhag 87% “premium” sirf drift tha. Zero se compare karte toh +0.79% “edge” lagta. (Is pe poora article alag se hai.)

Teesri cheez: “random se behtar hai?”

Drift ke baad ek aur test: same stock, random din. Kuch stocks naturally zyada volatile hote hain — unme kisi bhi din entry lo, bade moves aayenge. F&O ban list ke study mein ban se nikalne ke baad 10 din mein +1.30% market-adjusted return dikha, aur zero ke against bootstrap ne isko “significant” bataya. Lekin unhi stocks ke random non-ban dino pe same hold ka spread itna bada tha (standard deviation 8.4%) ki asli result ka z-score sirf +0.13 nikla. Matlab: noise.

Aapke liye ek 4-sawaal checklist

Jab bhi koi backtest, screenshot ya “90% accuracy” wala claim dekhein:

1. Base rate kya hai? Random din pe yeh outcome kitni baar hota hai?
2. Drift se compare kiya? Same dates pe market ne kya kiya?
3. Random entry se compare kiya? Same stocks, same exit rule, random din?
4. Kitne cheezein test ki? 40 ideas test karo toh 2 sirf luck se “significant” dikhenge. Kya multiple testing ke liye correct kiya?

Agar jawab “pata nahi” hai, toh claim ka matlab bhi “pata nahi” hai.

Yeh site isko kaise use karti hai

Har study mein aap N of M (kitne mein se kitne), window (kaunse saal), aur drift-adjusted number dekhenge. Jab koi cheez kaam nahi karti, hum saaf likhte hain ki dead hai — kyunki honest “nahi” bhi information hai. /learn/ section mein baaki concepts milenge.

The limitation

Drift benchmark bhi perfect nahi hai: equal-weight universe drift chhote stocks ko zyada weight deta hai, aur universe current listed stocks se bana hai (survivorship bias). Date-matched drift “market ne kya kiya” batata hai, “aapko kya karna chahiye” nahi. Base rates past describe karte hain; future predict nahi karte.

Methodology note. Base rate (50.3% of 10-day forward windows positive) from ~500 NSE stocks, daily data May 2023 – Oct 2026, entry next open, exit close +10 days. The four false edges and their honest re-tests: tietiy internal research, July 2026 (batch_a_alpha_control_2026-07-18, pead_honest_validation_2026-07-18). Pre-results figures: pre_earnings_premium_2026-07 (7,940 events, 187 F&O stocks, 2014–2026, cluster-robust by results season). Ban-list null: banlist_unwind_2026-07 (3,000 matched-stock random draws). All tests use date-matched equal-weight universe drift as the primary null.