Methodology
Every strategy in the book follows the same structured path from idea to deployment. This page walks through each stage, from testing ideas, validating, and eventually earning a spot in the portfolio. (Most don't survive)
Regimes
In-sample backtest (Jan 2019 - Dec 2024). Strategies were developed, optimized, and validated on this data. This is the period where overfitting risk is highest. Results here prove the concepts work historically but not necessarily that they'll work going forward.
Forward-test (Jan - Dec 2025). Strategy params were frozen. They ran on data they'd never seen during development. This is the closest analog to live trading before risking real capital.
Paper traded (Jan - Jun 2026). Real-time signal generation with no capital at risk. Same execution logic as live, same cost assumptions, but no money on the line. Used to verify the infrastructure end-to-end.
Live (From Jul 2026). Real capital on a live account.
Validation pipeline
Every strategy goes through a structured validation pipeline before deployment. Most ideas are rejected at some stage. (over 90% of strategies tested don't survive)
Out-of-sample testing. Strategies are trained on a rolling window, then tested on data they've never seen. Parameters are perturbed to check that edge doesn't disappear with small changes. Monthly returns are resampled thousands of times to estimate the robustness of outcomes. If a strategy only works on one specific window or one exact parameter set, it fails here.
Correlation filtering. New strategies are checked against the existing book. If pairwise correlation is too high, the strategy is rejected, even if its standalone backtest looks good.
Cost stress-testing. Costs are modeled at roughly +50% over the broker's average execution cost to account for slippage and variable market conditions, then stress-tested at 2x and 3x those levels to confirm the edge survives even under worst-case friction.
About drawdown & risk
Diversification across uncorrelated strategies. When one strategy is in drawdown, the others are typically doing something unrelated.
New strategies start small and only size up after proving themselves with real money over a meaningful sample. Allocations are determined by each strategy's risk profile, not equal weight, and reviewed periodically.
Per-strategy summary
4 uncorrelated strategies. Average pairwise correlation: +0.13. Range: +0.03 to +0.26.
Per-month strategy contributions are shown on individual report pages and in the History year sections. The backtesting engine is public on GitHub.
Lucas Guerin
lucasguerin.fr