Trading strategy and model backtesting

From a research idea to a historical test

Describe a rule-based strategy in chat or work directly in a Quant Lab notebook. Select the symbols, timeframe and date range relevant to the question. Review the strategy logic and configuration before interpreting the result.

For a trained model, open it in ML Models and use the regular Backtest flow. Supported custom models carry feature logic from training into historical prediction so the inputs are prepared consistently.

What to inspect in the result

• Returns and drawdowns. Examine the path of performance, including periods of loss, rather than only the final return.
• Trades. Inspect the trade list to understand which entries and exits produced the outcome.
• Configuration. Review the symbols, dates and portfolio assumptions associated with the run.
• Comparisons. Change a period or configuration and compare the resulting runs to see how sensitive the idea is.

Held-out evaluation and backtesting work together

Model training can compare a model against a simple baseline using held-out data. That comparison measures prediction quality for the chosen target. A model that improves a prediction metric may still produce a poor trading strategy, depending on how predictions become positions and the costs and assumptions used in the test.

Keep training and evaluation periods distinct when testing generalization. Inspect the data coverage and avoid features that use information unavailable at the time of a prediction.

Start with a small, inspectable example

Ask the assistant: “Train a ridge regression model using AAPL and MSFT daily data from the last two years to predict next-day returns. Compare it against a baseline on held-out data.” Review the metrics, save the model when you want to keep it, and open the ML Models backtest flow to evaluate a strategy using its predictions.

From historical tests to paper trading

Use paper trading to study a strategy's behavior as new data arrives. Historical and simulated results depend on their assumptions; they do not establish future profitability or reproduce every aspect of live execution.