Understand in-sample vs out-of-sample testing

What are In-Sample and Out-of-Sample?

WHY this split? Because strategies are built by trial-and-error. Every tweak you make to improve in-sample performance risks fitting random noise. Out-of-sample is your reality check: does the edge persist when the strategy faces new data?

WHAT happens without it? You get curve-fitting: a strategy with 15 parameters, 80% win-rate in-sample, that loses money the moment you trade it live. The classic quant graveyard.

HOW to split? Common approaches:

1. Chronological split: First 70% = IS, last 30% = OOS (respects time-series nature)
2. Walk-forward: Rolling IS/OOS windows (more robust, covered in advanced backtesting)
3. Leave-out segments: Hold out specific market regimes (e.g., 2008 crisis) as OOS

The Math: Why Overfitting Happens

Let's derive why more parameters → higher overfitting risk.

Setup: You have $N$ data points (e.g., daily returns). You fit a model with $p$ parameters.

Degrees of Freedom: The model "uses up" $p$ degrees of freedom to fit data. Remaining degrees of freedom for genuine signal: $\text{DoF} = N - p$.

Signal vs Noise: Assume returns = signal + noise, $r_t = s_t + \epsilon_t$, where $\epsilon_t \sim \mathcal{N}(0, \sigma^2)$ is random.

When you optimize $p$ parameters over $N$ points:

• Expected in-sample error: Your optimizer finds the best fit to $s_t + \epsilon_t$, including fitting some noise.
• Expected out-of-sample error: OS data has different noise $\epsilon_t'$, so the noise-fitted parts fail.

Example numbers:

• IS period: 252 days (1 year), $N = 252$
• Strategy has 10 parameters (moving average lengths, thresholds, stop-losses, etc.), $p = 10$
• Ratio: $p/N = 10/252 \approx 4\%$. Borderline acceptable.
• If $p = 50$, ratio = 20% → severe overfitting risk. You're fitting noise.

Worked Example: Moving Average Crossover

Strategy: Buy when 20-day MA crosses above 50-day MA. Sell when it crosses below.

Step 1: In-Sample Optimization

You have 5 years of SPY data (2017-2021, ~1260 days). You test all combinations:

• Short MA: 10, 15, 20, 25, 30 days
• Long MA: 40, 50, 60, 70, 80 days

That's $5 \times 5 = 25$ combinations (25 parameters tested).

Result: Best IS combo is (15, 70) with 18% annual return.

Why this step? You're searching for the parameter set that maximizes IS performance. This is standard optimization.

Step 2: Out-of-Sample Test

Reserve 2022-2023 (~500 days) as OOS. You never touched this data during optimization.

Run the (15, 70) strategy on OS data.

Result: OS return is 3% annualized.

Why the drop? The (15, 70) combo likely exploited random SPY fluctuations in 2017-2021 that didn't repeat in 2022-2023. Classic overfitting.

HOW to interpret?

Large gap = overfitting. The strategy has no real edge; it memorized IS noise.

What if OS was similar to IS? Say OS return = 14%. Gap = 4%. Still some overfitting, but the edge is real. A 20-30% degradation is typical (parameters are never perfectly tuned). A 80% collapse (18% → 3%) screams overfitting.

Worked Example 2: Mean Reversion with Multiple Filters

Strategy: Buy when RSI < 30 AND price < 200-day MA AND volume > 1.5× average.

Parameters:

• RSI threshold: $\theta_1 \in \{20, 25, 30, 35\}$
• MA length: $\theta_2 \in \{150, 200, 250\}$
• Volume multiplier: $\theta_3 \in \{1.2, 1.5, 2.0\}$

Total combinations: $4 \times 3 = 36$ parameter sets.

Data: 10 years AAPL (2013-2022), split 70/30 → IS: 2013-2019 (7 years), OOS: 2020-2022 (3 years).

Step 1: In-Sample Grid Search

Test all 36 combos. Best IS: RSI=25, MA=150, Vol=2.0 → 22% annual return, Sharpe 1.5.

Why this step? You're doing hyperparameter optimization (aka parameter tuning). Every backtest platform does this.

Step 2: Out-of-Sample Reality Check

Run (25, 150, 2.0) on OOS 2020-2022.

Result: 8% return, Sharpe 0.6.

Why the drop?

• 2020 had COVID crash (regime change)
• The (25, 150, 2.0) combo was tuned to 2013-2019's "normal" volatility
• Volume spike behavior changed in pandemic trading

HOW to fix?

1. Walk-forward testing (re-optimize every year using a rolling IS window)
2. Regularization: Penalize complex strategies (fewer parameters)
3. Ensemble: Average multiple parameter sets instead of picking the "best"

Key insight: Even with a 70/30 split, if OS includes a regime shift (2020 pandemic), your strategy will struggle. OS must be representative and unseen.

Common Mistakes (Steel-man)

Practical Implementation

HOW to split in Python (pandas):

import pandas as pd
 
df = pd.read_csv('AAPL_daily.csv', parse_dates=['Date'])
df = df.sort_values('Date')
 
split_date = '2020-01-01'
is_data = df['Date'] < split_date]  # In-sample
oos_data = df[df['Date'] >= split_date]  # Out-of-sample
 
print(f"IS: {len(is_data)} days, OOS: {len(oos_data)} days")

WHY chronological? Time-series data has temporal dependencies. Shuffling (like in ML) breaks this—you'd be training on "future" data to predict the "past." Chronological split respects causality.

WHAT about walk-forward?

# Pseudo-code for walk-forward
results = []
for start in range(0, len(df), 252):  # Roll every year
    is_window = df[start : start+756]  # 3 years IS
    oos_window = df[start+756 : start+1008]  # 1 year OOS
    
    best_params = optimize(strategy, is_window)
    oos_perf = backtest(strategy, oos_window, best_params)
    results.append(oos_perf)
 
avg_oos = np.mean(results)  # Aggregate OOS performance

WHY better? Multiple OOS periods reduce luck. If strategy works across 5 OS windows, it's more robust than one lucky OS period.

The 80/20 Core

20% of concepts that explain 80% of IS/OOS testing:

1. IS = where you optimize, OOS = where you validate. Never mix them.
2. Overfitting grows with $p/N$ ratio (parameters / data points). Keep $p/N < 5\%$.
3. Large IS/OOS performance gap (>40% drop) = overfitting. Discard the strategy.
4. Walk-forward > single split for robust validation (multiple OOS tests).

If you remember these four, you'll avoid 80% of backtesting disasters.

Diagram

What the diagram shows:

• Top panel: Equity curve in IS (blue) vs OOS (red). Notice IS is smooth (overfit), OOS is noisy (reality).
• Bottom panel: Parameter sensitivity. Optimal param in IS (green dot) gives poor OOS performance (red dot). Overfit.

Concept Map

Hinglish (regional understanding)