Understand walk-forward analysis — Hinglish notes

6.1.11 · Stock-Market › Algorithmic & Quant Trading

Walk-Forward Analysis Kya Hai?

Traditional Backtesting Kyun Fail Hoti Hai

Traditional backtesting: Pure historical dataset par parameters optimize karo → Usi data par test karo → Results report karo.

Teen fatal flaws:

Look-ahead bias: Optimize karte waqt tum "future jaante ho"
Overfitting: Parameters noise ke liye tune hote hain, signal ke liye nahi
No adaptation: Real traders periodically re-optimize karte hain; static parameters regime changes ko ignore karte hain

Example: Ek RSI(14) + Bollinger Band strategy jo 2015-2020 data par optimize hui ho, woh RSI threshold = 23.7 aur BB width = 2.13 select kar sakti hai kyunki woh exact values lucky trades pakadne mein kaam aayi thi. 2021 ke alag volatility regime mein, yeh parameters catastrophically fail kar jaate hain.

Walk-Forward Analysis Kaise Kaam Karta Hai

Mathematical Framework

Maano total dataset time $[T_{0}, T_{N}]$ span karta hai. Ise $K$ segments mein divide karo:

$Segment k : In-Sample [T_{k - 1}, T_{1} + L_{IS}] Out-of-Sample [T_{k - 1} + L_{IS}, T_{k - 1} + L_{IS} + L_{OOS}]$

Jahan:

$L_{IS}$ = In-sample window ki length (optimization period)
$L_{OOS}$ = Out-of-sample window ki length (testing period)
Agle segment ke liye $L_{step}$ se aage badhao (overlap ke liye $L_{OS}$ se kam ho sakta hai)

Yeh windows kyun?

IS window itni lambi honi chahiye ki market patterns capture ho sakein (jaise, 1-2 saal)
OOS window realistic re-optimization frequency represent karta hai (jaise, quarterly = 3 mahine)
Step size overlap control karta hai; chote steps = zyada OS periods lekin zyada correlated

Derivation: Walk-Forward Efficiency

Walk-Forward Efficiency (WFE) metric quantify karta hai ki tumhara IS performance OOS mein kitna survive karta hai:

Step 1: Sabhi segments mein average IS performance calculate karo

\overset{ˉ}{R}_{IS} = \frac{1}{K} k = 1 \sum K R_{IS, k}

jahan $R_{IS, k}$ segment $k$ ke in-sample period mein return (ya Sharpe ratio, profit factor, etc.) hai.

Step 2: Average OOS performance calculate karo

\overset{ˉ}{R}_{OOS} = \frac{1}{K} k = 1 \sum K R_{OOS, k}

Step 3: WFE ratio hai $WFE = \frac{R ˉ _{OOS}}{R ˉ _{IS}} \times 100%$

Yeh formula kyun?

Agar WFE ≈ 100%, toh OS performance IS se match karti hai → robust strategy
Agar WFE < 50%, tum overfit kar rahe ho
Agar WFE > 100%, toh OS mein lucky rehe (rare, suspicious)

First principles se derivation:

Trading non-stationary hai; ek period par optimize kiye gaye parameters generalize nahi ho sakte
WFE generalization measure karta hai: "Kya patterns jo maine dhoondhe woh unseen data mein bhi kaam karte hain?"
Ratio strategy ki aggressiveness ke liye normalize karta hai (ek high-return IS mein valid hone ke liye high-return OOS honi chahiye)

Worked Example 1: Simple Moving Average Crossover

Setup:

Data: 5 saal (2017-2021) ke daily SPY prices
Strategy: SMA crossover, short/long periods optimize karo
Configuration: IS = 1 saal, OOS = 6 mahine, step = 6 mahine

Execution:

| Segment | IS Period | OOS Period | Optimized Params | IS Return | OOS Return | |---------|-----------|------------------|-----------|----------| | 1 | 2017-2017 | 2018H1 | SMA(10,50) | +22% | +8% | | 2 | 2018-2018 | 2019H1 | SMA(15,60) | +18% | -3% | | 3 | 2019-2019 | 2020H1 | SMA(8,40) | +25% | -12% (COVID crash) | | 4 | 2020-2020 | 2021H1 | SMA(20,80) | +30% | +15% |

Yeh steps kyun?

Segment 1: 2017 ke bull market par optimize karo → Params continued bull (2018H1) mein acche kaam karte hain
Segment 2: 2018 volatility → Alag params, lekin 2019H1 whipsaw paise lose karta hai
Segment 3: 2019 steady growth → Aggressive params black-swan crash mein fail ho jaate hain (calm ke liye overfitted)
Segment 4: 2020 recovery → Smoother params rebound capture karte hain

WFE Calculate karo:

\overset{ˉ}{R}_{IS} = \frac{22 + 18 + 25 + 30}{4} = 23.75%

\overset{ˉ}{R}_{OOS} = \frac{8 - 3 - 12 + 15}{4} = 2%

WFE = \frac{2}{23.75} \times 100% \approx 8.4%

Verdict: Bahut bura WFE! Strategy overfit karti hai. High IS returns OS mein translate nahi hote.

## Worked Example 2: Mean-Reversion with RSI

Setup:

Data: 3 saal (2019-2021) ke AAPL 1-hour bars
Strategy: Buy karo jab RSI(period=P) < threshold T, +2% par sell karo
Configuration: IS = 6 mahine, OOS = 2 mahine, step = 2 mahine

Execution:

| Segment | IS Period | OS Period | Optimal (P, T) | IS Sharpe | OOS Sharpe | |---------|-----------|------------|-------------|----------| | 1 | 2019H1 | 2019 Jul-Aug | (14, 28) | 1.8 | 1.5 | | 2 | 2019 Jul-Dec | 2020 Jan-Feb | (12, 25) | 2.1 | 0.9 | | 3 | 2020 Jan-Jun | 2020 Jul-Aug | (10, 30) | 1.6 | 1.2 | | 4 | 2020 Jul-Dec | 2021 Jan-Feb | (16, 32) | 2.3 | 1.9 | | 5 | 2021 Jan-Jun | 2021 Jul-Aug | (14, 26) | 1.9 | 1.6 |

Yeh results kyun?

Segment 2: COVID uncertainty → Market structure toot jaata hai, OS suffer karta hai
Segments 4-5: Stable regime → Parameters consistent rehte hain, OOS decent hai

WFE Calculate karo (Sharpe use karke):

\overset{ˉ}{S}_{IS} = \frac{1.8 + 2.1 + 1.6 + 2.3 + 1.9}{5} = 1.94

\overset{ˉ}{S}_{OOS} = \frac{1.5 + 0.9 + 1.2 + 1.9 + 1.6}{5} = 1.42

WFE = \frac{1.42}{1.94} \times 100% \approx 73%

\text{Rolling Window}
$IS_{k} = [T_{k - 1}, T_{k - 1} + L_{IS}] (fixed length, moves forward)$
Anchored Window:
$IS_{k} = [T_{0}, T_{k - 1} + L_{IS}] (start se grow karta hai)$
Yeh difference kyun matter karta hai:

Rolling: Recent regimes ke liye adapt karta hai, purana data bhool jaata hai. Non-stationary markets (crypto, forex) ke liye better.

Anchored: Saara historical data use karta hai, zyada stable parameters. Long-term strategies ke liye better.

Choice ki Derivation: Agar market $M_{t}$ non-stationary hai (distribution $P_{t} (r)$ change hoti hai), toh:
$E [error] \propto \int_{T_{0}}^{T_{k}} ∣ P_{t} (r) - P_{current} (r) ∣ d t$

Anchored windows ka high integral hota hai (purana irrelevant data include karta hai), rolling windows ka low integral hota hai (sirf recent data).

## Common Mistakes

Common mistake Mistake 1: IS aur OOS ke liye Same Data Use Karna

Galat idea: "Main 2015-2020 par optimize karunga, phir 2015-2020 par 'test' karunga dekhne ke liye ki parameters acche hain." Kyun sahi lagta hai: Tum check kar rahe ho ki optimization kaam kiya.

Fix: Yeh circular hai! Tum usi data par test kar rahe ho jo optimizer ne pehle se dekha hai. OS chronologically IS ke BAAD hona chahiye. Optimizer OS data peek nahi kar sakta.

Steel-man: Yeh mistake isliye hoti hai kyunki log "validation" (optimizer logic check karna) aur "generalization testing" (real-world viability check karna) mein confuse ho jaate hain. WFA baad wala karta hai.

Common mistake Mistake 2: Best WFA Configuration Cherry-Pick Karna

Galat idea: "Maine IS=1yr/OS=3mo, IS=2yr/OOS=6mo, IS=6mo/OOS=1mo try kiya... Main woh report karunga jiska WFE best hai."

Kyun sahi lagta hai: Tum "optimal" testing setup dhundh rahe ho.

Fix: Ab tumne meta-parameters ko overfit kar liya! WFA window sizes trading constraints ke basis par choose hone chahiye (tum realistically kitni baar re-optimize kar sakte ho?) WFA run karne se PEHLE. Typical: IS=1-2 saal, OOS=3-6 mahine daily strategies ke liye.

Steel-man: Temptation real hai—better results investors ko better lagte hain. Lekin tum sirf overfitting ko ek level upar le ja rahe ho. Discipline zaroori hai.

Common mistake Mistake 3: OOS Periods mein Transaction Costs Ignore Karna

Galat idea: "Mera WFA 50% returns ke saath 80% WFE dikhata hai... ship kar do!"

Kyun sahi lagta hai: Tumne returns par focus kiya, jo main metric hai.

Fix: High-frequency re-optimization (short OS periods) parameter churn kausi karti hai—tumhare position sizes aur signals har mahine change hote hain, commissions aur slippage rackup karte hain. Calculate karo:

$R_{net} = R_{gross} - N_{trades} \times C_{per-trade} - Spread \times N_{trades}$
jahan $N_{trades}$ mein woh trades bhi shamil hain jo OOS periods ke beech parameter changes se trigger hoti hain.

Steel-man: Backtests aksar "zero-cost" assumptions use karte hain. Real WFA ko har OS transition par realistic costs model karni chahiye.

Active Recall Flashcards

#flashcards/stock-market

Walk-forward analysis ka purpose kya hai? :: Walk-forward analysis trading strategies ko validate karta hai in-sample data par repeatedly optimize karke aur out-of-sample data par chronological order mein test karke, real-world adaptive trading simulate karta hai aur overfitting detect karta hai.

Walk-forward segment mein teen periods kaunse hote hain?

1) Parameter optimization ke liye In-Sample (IS) period, 2) Optimized parameters ko unseen data par test karne ke liye Out-of-Sample (OOS) period, 3) Naye IS aur OOS windows ke saath agle segment mein Step/roll forward.

Walk-Forward Efficiency (WFE) kya measure karta hai?

WFE = (Average OS Performance / Average IS Performance) × 100%, yeh measure karta hai ki in-sample performance kitni out-of-sample results mein translate hoti hai. High WFE (≈100%) robustness indicate karta hai; low WFE (<50%) overfitting indicate karta hai.

WFA mein OS data chronologically IS data ke baad kyun hona chahiye?

Look-ahead bias se bachne ke liye. Real trading mein, tum sirf past data par optimize kar sakte ho aur future mein trade kar sakte ho. Agar OS, IS se overlap kare ya pehle aaye, toh optimizer ne "future dekh liya," jo test ko invalid kar deta hai.

WFA mein anchored aur rolling windows mein kya fark hai?

Rolling windows fixed-length IS periods use karte hain jo aage slide karte hain (jaise, hamesha last 1 saal), regime changes ke liye jaldi adapt karte hain. Anchored windows ek fixed start date se expand hote hain (jaise, 2015 se present tak ka saara data), zyada data use karte hain lekin kam adaptive hote hain.

4 WFA segments wali strategy ka WFE kaise calculate karoge jisme IS returns [20%, 25%, 18%, 22%] aur OOS returns [12%, 8%, 15%, 10%] hain?

Average IS = (20+25+18+22)/4 = 21.25%, Average OOS = (12+8+15+10)/4 = 11.25%, WFE = (11.25/21.25)×100% ≈ 53%, jo moderate overfitting indicate karta hai.

Walk-forward analysis mein parameter churn kya cause karta hai?

Jab OS periods short hote hain (jaise, monthly re-optimize karo), optimal parameters segments ke beech frequently change hote hain, strategy ko positions aksar alter karne par majboor karte hain aur extra transaction costs incur hote hain jo real returns erode kar dete hain.

WFE > 100% suspicious kyun hai?

Iska matlab hai ki OS performance IS performance se zyada thi, jo suggest karta hai ya toh: 1) OOS periods mein extreme luck, 2) Data snooping (OOS data IS mein leak hua), ya 3) Insufficient IS optimization (undertrained parameters). Yeh rarely true strategy superiority indicate karta hai.

Daily equity strategies ke liye typical IS/OOS window configuration kya hai?

IS = 1-2 saal (252-504 trading days) market cycles capture karne ke liye, OOS = 3-6 mahine (63-126 days) realistic re-optimization frequency simulate karne ke liye, step = non-overlapping tests ke liye OS length.

WFA curve-fitting kaise detect karta hai?

Curve-fitted strategies IS data ke noise ke liye optimize hoti hain. Kyunki noise OOS mein repeat nahi hota, yeh strategies large IS-OOS performance divergence dikhati hain, jiske result mein low WFE aur segments mein inconsistent OS returns aate hain.

Connections

6.1.1-Backtesting-Fundamentals: WFA ek advanced backtesting methodology hai jo look-ahead bias address karta hai
6.1.8-Overfitting-and-Data-Snooping: WFA directly overfitting se ladta hai unseen data par test karke
6.1.9-Parameter-Optimization: WFA mein IS period woh jagah hai jahan parameter optimization hota hai
6.1.12-Monte-Carlo-Simulation: WFA ke saath combine kar sakte hain—OOS periods par MC run karo confidence intervals paane ke liye
6.2.5-Strategy-Robustness-Testing: WFA live deployment se pehle ek key robustness test hai
5.3.7-Sharpe-Ratio: Aksar WFE calculations mein raw returns ki jagah performance metric ke roop mein use hota hai
4.5.3-Regime-Detection: WFA mein rolling windows strategies ko regime changes ke liye adapt karne mein help karte hain

Recall Ek 12-Saal-Ke-Bacche Ko Explain Karo

Imagine karo tum basketball mein free throws practice kar rahe ho. Buri practice: Tum 100 baar shoot karte ho, apni technique adjust karte ho un exact 100 shots ko perfect banane ke liye (shayad hawa ek certain taraf chal rahi thi, ya rim thoda bent tha). Phir tum kehte ho "Main 90% shooter hun!" Lekin kal ek real game mein, tum sab kuch miss kar dete ho kyunki conditions change ho gayi.

Walk-forward practice: Tum 20 baar shoot karte ho aur apni technique adjust karte ho. Phir tum 10 NEW shots bina adjust kiye shoot karte ho—kya tumhari technique kaam ki? Phir tum 20 aur practice karte ho aur 10 aur test karte ho. Har test aise shots use karta hai jo tumne abhi tak practice nahi ki, isliye tum exact conditions yaad karke "cheat" nahi kar sakte.

Agar tumhare "test shots" almost utni hi baari jaate hain jitni tumhare "practice shots," toh tumhari technique REAL hai. Agar test shots fail ho jaate hain, tum sirf memorize kar rahe the, seekh nahi rahe the. Walk-forward analysis trading strategies ke liye "real game" test hai.