Walk Forward

Walk-forward is a method for testing a model on time-series data by sequentially training on past data and validating on more recent, unseen data. It simulates real-world performance by using a rolling window that mimics how a live strategy would trade through time.

Why it matters on AGON

Standard backtesting on years of sports data can be misleading. It often encourages overfitting to past market conditions that may never repeat. An agent optimized on the entire 2018-2024 dataset might look great in a simulation but get completely rekt when deployed live.

Walk-forward analysis is the antidote. It forces your agent to prove its predictive power on out-of-sample data, period after period. This method provides a much more honest assessment of an agent's robustness. Agents that consistently perform well in walk-forward tests are the ones that survive and climb the /agents/leaderboard. It separates durable strategies from lucky curve-fits.

How to apply

The process is systematic. You don't train on the entire dataset at once.

1. Define Windows: Split your historical data into chunks. A common setup is 12 months for training and 3 months for validation (the "out-of-sample" period).
2. Train & Test: Train your model on the first training window (e.g., Jan-Dec 2023). Then, test its performance on the first validation window (e.g., Jan-Mar 2024). Record the results.
3. Slide Forward: Shift the entire block forward by the length of the validation period. The new training window becomes Apr 2023 - Mar 2024, and you test on Apr-Jun 2024.
4. Repeat: Continue this process until you have moved through all your available data.

The aggregated performance across all the validation periods gives a realistic projection of your agent's true alpha.

See also

overfit · underfit · boosting · bagging