Learn look-ahead bias avoidance — Hinglish notes

6.2.4 · Stock-Market › Backtesting Frameworks

What is Look-Ahead Bias?

Definition Look-Ahead Bias

Look-ahead bias tab hota hai jab ek trading strategy backtesting mein aisi information use karti hai jo simulated trade decision ke waqt available nahi hoti. Isse performance metrics artificially inflate ho jaate hain jo real-world tradability ko reflect nahi karte.

Do core forms:

Data look-ahead: Revised/restated data use karna originally reported values ki jagah, ya koi value uski actual availability date se pehle use karna
Temporal look-ahead: Aaj ka close price use karke aaj ka signal generate karna (aisi info par act karna jo abhi exist hi nahi karti)

Related lekin alag: Survivorship bias (sirf un stocks par test karna jo aaj tak survive kar gaye) ek data-selection bias hai, strictly look-ahead issue nahi—haalaanki dono aksar saath aate hain. Inhe conceptually alag rakho: look-ahead time of availability ke baare mein hai, survivorship is baare mein hai ki kaun si entities include ki gayi.

Why Look-Ahead Bias Happens

Root causes:

Data timestamp confusion: Financial data ke multiple timestamps hote hain (report date, filing date, announcement time, database update time). Galat wala use karne se bias create hota hai.
Point-in-time data unavailability: Zyaadaatar data vendors current/restated data dete hain, na ki woh exact values jo historically available thi.
Invisible corporate actions: Stock splits, delistings, aur bankruptcies ko cleaned datasets mein retroactively handle kiya jaata hai.
Signal-execution misalignment: Time T par signal generate karna lekin execution bhi time T par assume karna (impossible hai—execution ke liye kam se kam T+1 chahiye).

How to Detect Look-Ahead Bias

Method 1: Forward-Walk Verification

Derivation from first principles:

Maano $S_{t}$ = time $t$ par strategy signal, $I_{t}$ = time $t$ par available information set.

Causal constraint: $S_{t}$ sirf $I_{t}$ ka function hona chahiye: $S_{t} = f (I_{t}) where I_{t} \cap I_{> t} = \emptyset$

Time $t$ par information set future information se strictly disjoint hai.

Test: Har decision timestamp $t$ ke liye:

Sirf $t$ ke timestamp wale data use karke $I_{t}$ reconstruct karo
$S_{t}$ dobara generate karo
Backtest signal se compare karo

Agar $S_{t}^{backtest} \neq = f (I_{t})$ , toh look-ahead bias hai.

Ye step kyun? Hum verify kar rahe hain ki causal arrow sirf past ki taraf point karta hai.

Method 2: Reasonable Performance Bounds

Derivation from first principles:

Maano hum $N$ periods mein strategy ka Sharpe ratio $S R_{s}$ estimate karte hain aur use benchmark $S R_{m}$ se compare karte hain. $N$ observations par ek Sharpe estimate ki standard error approximately hoti hai:

$σ_{S R} \approx \frac{1}{N}$

(i.i.d. returns ke liye; ye already $N$ badhne par shrink hoti hai). Ye poochhne ke liye ki "kya strategy ka Sharpe benchmark se significantly upar hai?", hum ek standard z-statistic banate hain difference ko uski standard error se divide karke:

$z = \frac{S R _{s} - S R _{m}}{σ _{S R}}$

Flag condition: Hum bias ke liye investigate karte hain jab

$z = \frac{S R _{s} - S R _{m}}{σ _{S R}} > 2$

Extra $N$ kyun nahi? Kyunki $σ_{S R}$ mein pehle se $1/ N$ scaling hai. Ek aur $N$ multiply karna sample size ko double-count karega aur test galat ho jaayega. Correct significance test simply $(S R_{s} - S R_{m}) / σ_{S R} > 2$ hai (roughly 95% confidence).

Ye step kyun? Bahut zyaada Sharpe ratios (daily strategies ke liye >3, intraday ke liye >5) huge $z$ -values produce karte hain—ek statistical red flag. Market competitive hai; persistent excess returns ke liye luck se aage explanation chahiye. Unrealistically achhe results aksar data leakage indicate karte hain, genius nahi.

Avoidance Strategies

Strategy 1: Timestamp Discipline

Example implementation:

# WRONG - look-ahead bias
df['signal'] = (df['close'] > df['ma_20'])  # Uses today's close for today's signal
 
# RIGHT - proper timing
df['signal'] = (df['close'].shift(1) > df['ma_20'].shift(1))  # Yesterday's data for today

Ye step kyun? Shifting ensure karta hai ki signal sirf woh information use kare jo trading decision se pehle exist karti thi.

Strategy 2: Point-in-Time Database

Concept: Har value ko uski availability timestamp ke saath store karo, na ki sirf reference timestamp ke saath.

Schema design:

symbol	metric	value	ref_date	avail_date
AAPL	EPS	1.20	2024-12-31	2025-01-28
AAPL	EPS	1.24	2024-12-31	2025-02-15

Query rule:

SELECT value 
FROM fundamentals 
WHERE symbol = 'AAPL' 
  AND ref_date = '2024-12-31'
  AND avail_date <= :backtest_date  -- Key: as-of filter
ORDER BY avail_date DESC 
LIMIT 1;

Ye step kyun? Ye us exact value ko capture karta hai jo ek trader kisi bhi historical date par dekhta, revisions samait.

Strategy 3: Trade-at-Open Protocol

Rule: Kal ke close tak ke data use karke signals generate karo, kal ke open par execute karo.

Timeline: $t_{data cutoff} = Day T ’s close \Rightarrow t_{signal} = Day T + 1 ’s open$

Kyun? Yeh conservative approach guarantee karta hai ki koi intraday look-ahead nahi hai aur overnight processing time ko account karta hai.

Worked example Worked Example: Momentum Strategy

Scenario: AAPL par 20-day momentum breakout strategy test karo, 2024-01 se 2024-12-31 tak.

Step 1: Signal define karo. Breakout ka matlab hai "kal ka close kal se pehle ke 20 din ke highest close se zyaada tha": $S_{t} = 1 {P_{t - 1} > max_{i = 2}^{21} P_{t - i}}$ jahan $P_{t}$ = day $t$ par close price, $1$ = indicator function.

Ye step kyun? Candidate price $P_{t - 1}$ ko aisi window ke against compare karna chahiye jo khud $P_{t - 1}$ ko exclude kare—yaani $P_{t - 2}, P_{t - 3}, \dots, P_{t - 21}$ . Warna tum $P_{t - 1}$ ko ek max se compare karte ho jo pehle se $P_{t - 1}$ ko contain karta hai, aur strict-greater test kabhi theek se fire nahi kar sakta.

Step 2: Timestamps verify karo

Signal calculate kiya: Day $t - 1$ ke market close ke baad
Execution: Day $t + 1$ ka market open (open chahiye, jo signal ke baad hai)
Data used: Day $t - 1$ tak ke saare prices

Check: Kya hum day $t - 1$ close hone ke baad $max_{i = 2}^{21} P_{t - i}$ compute kar sakte hain? Haan—saare 20 historical prices ( $P_{t - 2}$ se $P_{t - 21}$ tak) available hain.

Ye step kyun? Explicit verification ki har input strictly past mein hai.

Step 3: Code implementation

# Window is the 20 closes BEFORE yesterday's close:
# shift(2) drops today (t) and yesterday (t-1), rolling(20) covers t-2..t-21
df['max_20'] = df['close'].shift(2).rolling(20).max()
 
# Signal: yesterday's close (shift(1)) breaks above that window
df['signal'] = (df['close'].shift(1) > df['max_20']).astype(int)
 
# Execution: enter at the OPEN following the signal, hold to next open.
# Signal known after close of t-1 -> trade open-to-open of days t..t+1
df['open_ret'] = df['open'].shift(-1) / df['open'] - 1   # open[t+1]/open[t]-1
df['returns'] = df['signal'] * df['open_ret']

shift(2) window ke liye kyun? Ye today ( $t$ ) aur candidate day ( $t - 1$ ) dono ko remove karta hai taaki max exactly $P_{t - 2} \dots P_{t - 21}$ span kare—formula se match karta hai.

Open-to-open returns kyun? Kyunki hum open par execute karte hain, realized P&L woh change hai jo us open se next open tak hota hai jahan hum exit karte hain—na close-to-close return, jo hum kabhi capture nahi kar sakte the.

Result: Sharpe ratio ≈ 0.8 (realistic), max drawdown ≈ 18%, koi look-ahead bias detect nahi hua.

Worked example Worked Example: Earnings Surprise Strategy

Scenario: Un stocks ko buy karo jo earnings beat karte hain >5% se, jo miss karte hain >5% se unhe sell karo.

Naive approach (WRONG):

df['surprise'] = (df['eps_actual'] - df['eps_estimate']) / df['eps_estimate']
df['signal'] = np.where(df['surprise'] > 0.05, 1, -1)

Problem: eps_actual kab available hoti hai? Reference date par nahi—announcement date par!

Ye kyun fail hota hai: Agar AAPL ki Q4 2024 earnings (ref_date = 2024-12-31) 2025-01-28 ko announce hoti hain, toh unhe 2024-12-31 ko use karna time travel hai.

Correct approach:

df['announce_date'] = df['earnings_announce_date']  # Explicitly tracked
df['avail_next_day'] = df['announce_date'] + pd.Timedelta(days=1)  # T+1 trade
df_trades = df[df.index >= df['avail_next_day']]  # Only trade after announcement

Ye step kyun? Hum actual announcement date par merge karte hain, na ki fiscal period end date par.

Result: Sharpe ratio 2.4 (biased) se gir ke 0.6 (realistic) ho jaata hai. Bias ~15% annual return ke barabar tha jo kabhi exist hi nahi karta tha.

Common Mistakes

Common mistake Mistake 3: Confusing Survivorship Bias With Look-Ahead Bias

Galat idea: "Survivorship sirf ek aur type ka look-ahead bias hai, toh ek fix karna dono fix kar deta hai."

Ye sahi kyun lagta hai: Dono backtests inflate karte hain aur dono mein "future ke baare mein kuch jaanna" involved hai.

Reality: Ye alag hain. Look-ahead bias time of information availability ke baare mein hai (data ko exist karne se pehle use karna). Survivorship bias ek data-selection bias hai (delisted/dead entities ko universe se exclude karna). Tum look-ahead bilkul eliminate kar sakte ho aur phir bhi survivorship bias ho sakta hai, aur vice versa.

The fix: Inhe alag-alag address karo—look-ahead ke liye point-in-time timestamps use karo, aur survivorship ke liye historically accurate constituents wala survivorship-bias-free universe use karo.

Impact: Inhe ek samajhne se tum sirf aadhi problem "fix" karte ho.

Prevention Checklist

Koi bhi backtest chalane se pehle:

☐ Har data field ka ek explicit availability timestamp hai
☐ Signal generation time t par decision ke liye sirf data[t-1] use karta hai
☐ Rolling windows candidate bar khud ko exclude karti hain
☐ Returns execution timing se match karte hain (open par trade karne par open-to-open)
☐ Execution realistic delay assume karta hai (open, ya close+1 day)
☐ Corporate actions unki effective date par apply hote hain, retroactively nahi
☐ Index constituents backtest date ke anusaar verified hain (survivorship alag handle)
☐ Results "reasonable Sharpe" test pass karte hain ( $z = (S R_{s} - S R_{m}) / σ_{S R} < 2$ ; SR < 3 daily)
☐ 10 random dates par forward walk-through kiya gaya hai

Recall 12-Saal-Ke-Bacche Ko Explain Karo

Socho tum ek video game khel rahe ho jahan tumhe guess karna hai ki soccer match mein kaun jeeta. Look-ahead bias aisa hai jaise tum game dekh lo, dekh lo kaun jita, aur phir time mein peeche jaake winner "predict" karo. Bilkul sahi rahoге 100% time—tumhe pehle se pata tha!

Stock trading mein, look-ahead bias tab hota hai jab tum ek strategy aisi information use karke test karte ho jo tumhare paas actually hoti nahi. Jaise aaj ki final stock price use karke decide karna ki aaj subah khareedna hai ya nahi. Yeh cheating hai! Jab tum real mein try karte ho (bina time travel ke), teri strategy fail ho jaati hai.

Isse bachne ke liye: pretend karo ki tum actually past mein jee rahe ho. Sirf woh news, prices, aur reports use karo jo har decision se PEHLE exist karti thi. Yeh game fair khelne jaisa hai—ending peek mat karo!

Connections

6.1.01-Understand-backtest-fundamentals - Foundational backtest design
6.2.03-Handle-survivorship-bias - Related lekin distinct data-selection bias
6.2.05-Account-for-transaction-costs - Realistic execution modeling
6.3.02-Implement-walk-forward-analysis - Validation method jo look-ahead detect karne mein help karta hai
7.1.04-Handle-asof-merges-in-pandas - Point-in-time data ki technical implementation

#flashcards/stock-market

Look-ahead bias kya hai?

Backtest mein aisi information use karna jo simulated trade decision ke waqt available nahi hoti thi, performance artificially inflate karta hai.

Look-ahead bias aur survivorship bias mein kya fark hai?

Look-ahead bias time of information availability ke baare mein hai (data ko exist karne se pehle use karna); survivorship bias ek data-selection bias hai (delisted/dead entities ko exclude karna). Ye alag hain haalaanki aksar saath aate hain.

Adjusted stock prices ke saath look-ahead bias kyun hota hai?

Adjusted prices splits aur dividends ko saare historical data par retroactively apply karti hain, lekin traders past decisions karte waqt future corporate actions ke baare mein nahi jaante the.

As-of data rule kya hai?

Time t par kisi bhi calculation ke liye sirf aisa data use karo jiska data_timestamp ≤ t - delay ho, jahan delay reporting, processing, aur market delays account karta hai.

Look-ahead bias se bachne ke liye kaun sa timeline follow karna chahiye?

Kal ke close tak ke data use karke signals generate karo (day T), next open par execute karo (day T+1), kam se kam ek session lag ensure karke.

Aaj ki close price par trade kyun nahi kar sakte?

Close price tum sirf market band hone ke baad jaante ho, toh trading opportunity pehle hi ja chuki hoti hai. Agli trading session tak wait karna padta hai.

Suspiciously high Sharpe ke liye correct significance test kya hai?

z = (SR_s − SR_m) / σ_SR > 2, jahan σ_SR pehle se 1/√N scale karta hai. Koi extra √N factor apply nahi hota.

(SR_s − SR_m) > 2√N·σ_SR likhna galat kyun hai?

Kyunki σ_SR mein pehle se 1/√N scaling hai; ek aur √N multiply karna sample size ko double-count karta hai aur test tod deta hai.

20-day breakout mein kal ke close P_{t-1} use karke max kis window par span karna chahiye?

P_{t-2} se P_{t-21} tak (yaani shift(2).rolling(20)), candidate P_{t-1} khud ko exclude karke.

Agar tum open par execute karte ho, backtest mein kaun sa return use karna chahiye?

Open-to-open (open[t+1]/open[t] − 1) ya open-to-close entry timing match karte hue—kabhi close-to-close nahi, jo tradeable nahi hai.

Point-in-time database kya hai?

Ek database jo har data value ko uski availability timestamp ke saath store karta hai (sirf reference timestamp nahi), taaki queries ki ja sakein ki "koi value historical date X par kaisi dikhti thi."

Kaun sa Sharpe ratio threshold possible look-ahead bias suggest karta hai?

Daily strategies jinka Sharpe > 3 ho ya intraday strategies jinka Sharpe > 5 ho, ye red flags hain jo bias ya data leakage ki investigation maangti hain.