Backtest Engine — ^GSPC

Generic Event Backtester · Grid Search · Walk-Forward · AI Relevance

What does this page show? — The Backtest Engine tests seasonal trading strategies on historical data. Choose an event (FOMC, OPEX, holiday, calendar day…), define entry/exit offsets and examine the equity curve, win rate and significance. 4 tabs: Single backtest, grid optimisation (all parameter combinations), walk-forward analysis (out-of-sample) and event relevance ranking. Sidebar: Ticker, period, event type, offsets, stop-loss and indicator filters — all combinable.

Backtest Parameters

Days BEFORE Event (Entry) 3

Days AFTER Event (Exit) 3

Grid-Search Parameters

days_before — Min ↔ Max

↔

days_after — Min ↔ Max

↔

Objective

Walk-Forward Configuration

Optimised on in-sample, tested on out-of-sample → realistic performance without overfitting.

Number of Folds 5

In-Sample Ratio 0.70

Objective

Relevance Calculation

Which events are statistically significant? t-test + effect size (Cohen's d) + win rate → relevance score.

Days BEFORE Event 3

Days AFTER Event 3

📖 How to Read the Relevance Score

The relevance score combines three dimensions into an overall value between 0 and 1:

Significance (50%): 1 - p-value
Win Rate (30%): Share of trades > 0
Effect Size (20%): Cohen's d, capped at 1.0

t-statistic: Signal strength vs. null. |t| > 2 = robust effect. p-value: Probability of chance. p < 0.05 = significant. Cohen's d: Effect size. d > 0.5 = medium, d > 0.8 = large.

Methodology

Backtest Logic

Each event date defines a trade: entry = event_date − days_before trading days, exit = event_date + days_after trading days. Minimum gap between trades = 5 days (prevents overlap).

Metrics

Total Return: cumulative equity from 100 (compound).
Sharpe Ratio: Avg Return / Std × √(trades per year). Risk-free = 0.
Calmar Ratio: Annualized Return / Max Drawdown.
Max Drawdown: Maximum equity decline in %.
Profit Factor: Σ Wins / |Σ Losses|.

Stop-Loss

Stop is checked close-based (no intraday low/high in DB). Fixed: stop at entry × (1 - stop%). Trailing: stop at high_watermark × (1 - stop%), watermark tracks highest close.

Walk-Forward

Expanding window: optimise on the past, test on the future. Each fold runs a grid search on in-sample years, then applies the best parameters to out-of-sample years. The aggregated OOS performance is the most honest estimate of live performance.

Event Relevance Score

One t-test per event (H0: mean = 0). Score = 50% significance (1−p) + 30% win rate + 20% effect size (Cohen's d, capped at 1.0).