ML Model Training
Train custom market models in chat or use preset workflows. Compare held-out metrics with a baseline, save models and backtest them in Sigmo.
Model training is usually where retail research stops: the data is scattered, the laptop is too small and the pipeline takes weeks to build. Sigmo runs the whole loop as a product. Pick a universe, choose features and a label, train on our compute, and get a model with its backtest, metrics and configuration recorded.
Train a custom model in plain language
Describe the target, symbols, data window and algorithm you want. For example: “Train a ridge regression model on AAPL and MSFT daily data from the last two years to predict next-day returns. Compare it with a baseline on held-out data.”
The assistant generates a training notebook and prepares reusable feature logic for supported market-model workflows. You can inspect and refine the code and configuration. Training runs in isolated compute; model and feature checks help determine whether the result can be used in the platform's prediction and backtesting pipeline.
Compare the model with a baseline
Review the held-out metrics alongside the baseline comparison. The model card shows the score, whether the comparison passed and the improvement relative to the selected baseline. Adjusting the comparison baseline changes that comparison; it does not retrain the model or change its measured score. Passing a baseline is evidence about that evaluation, not a guarantee of profitable trading.
Save only what you want to keep
A chat-trained model is a preview until you click Save model. Once saved, find it in ML Models and open the regular backtesting flow. Supported custom feature logic prepares the historical inputs, so you do not need to assemble a features JSON file by hand. Save a reusable training preset when you want to run the same logic with another supported configuration.
Universe, features, label
- Universe presets. S&P 500, Russell 2000 and Russell 3000 reconstructed from historical membership, or your own list, so the training set is what was actually investable at the time.
- Feature library. Price and volume features across timeframes, fundamentals with point-in-time dates, short interest, options data, institutional and insider positioning, alternative data and macro series, all aligned to the same calendar.
- Labels and horizons. Forward returns, ranks and classification targets over the horizon you choose.
- Timeframes. Daily models for swing and position work; intraday models for shorter horizons.
Training runs that are reproducible
Runs are queued on Sigmo's training cluster, not your machine. Each run records its universe, data window, feature set, label and metrics, and a run with the same configuration joins the existing result instead of training twice.
Backtest before you believe it
Compatible models can be backtested over historical data with the portfolio construction you choose. You get returns, drawdowns, the trade list and a tearsheet, together with the assumptions the test made.
Put the model to work
Attach a model to the agent's auto-trade strategy on a paper account, use it as a ranking system in the screener, or call it from the SDK. Your models sit next to Sigmo's own in the model list, with their backtest results one click away.
Included compute
Pro plans include daily training runs; Max adds monthly compute credits for larger universes and longer histories. See the pricing page.
See it on your own watchlist
Sigmo is free to start. Open the app, ask one question about a ticker you follow, and take it from there.