What is walk-forward validation for stock models?

Walk-forward validation trains on an earlier window, tests on the next unseen window, then rolls forward. TradeWell requires this out-of-sample pattern for research architectures before any promotion to live TW Quant Score rankings.

Is walk-forward the same as NDCG@30?

No. Walk-forward out-of-sample testing is a research gate for challenger architectures. Live rankers on the public board are evaluated on NDCG@30 (NDCG@20 when the evaluated list is shorter than 30).

Does clearing walk-forward auto-promote a model?

No. Promotion is operator-only. A research name is not live because it appears in a paper, a Slack card, or a challenger job.

Is TradeWell investment advice?

No. TradeWell publishes research scores and forecasts, not investment advice. Nothing on the site or in the app is a recommendation to buy or sell any security.

Walk-forward validation for AI stock rankings

Searching for walk-forward validation on stock models, or how an AI stock predictor proves itself out of sample? This page explains TradeWell’s research gate: train on an earlier window, test on the next unseen window, roll forward — and never treat a paper result as a live TW Quant Score.

TW Quant Score is research, not investment advice. It is not a guarantee of future performance and does not tell you what to buy or sell. Past model performance does not guarantee future results. TradeWell does not publish fabricated accuracy percentages as product claims.

What walk-forward means here

A single backtest on the same window used for fitting is not enough. Walk-forward out-of-sample (OOS) validation moves the train/test cut forward through time so each test slice is data the model had not used for fitting that fold. TradeWell requires that pattern for research architectures before any challenger can be considered for the public board.

Train on an earlier market window for a stated horizon.
Test on the next unseen window for that same horizon.
Roll the windows forward and repeat — do not peek.
Refuse fabricated win rates or IC figures dressed up as product claims.

Research gate vs live board

Walk-forward OOS is the research gate. The scores you see live are a separate surface: live rankers are evaluated on NDCG@30 (or NDCG@20 when the evaluated list is shorter than 30). Clearing a walk-forward job does not auto-promote a model. Promotion is operator-only. For the ranking metric itself, see how TradeWell accuracy is measured.

Why this matters for ten horizons

TradeWell publishes a separate TW Quant Score, expected return, and model-implied price target on each of ten horizons (1D through 5Y). Scores are not normalized across windows. A fold that looks fine on 1W is not evidence for 1Y. Walk-forward and live evaluation stay horizon-specific — see the ten prediction horizons explainer and the TW Quant Score page.

What this page will not claim

We will not invent a win rate, hit rate, or IC percentage for marketing. If a fold is warming or a metric is missing, we say so. Challenger model names are not “live” unless the methodology page says they currently serve the public board. For product shape (score + return + target per horizon), see the AI stock predictor page.

Walk-forward validation — FAQ

What is walk-forward validation for stock models?: Train on an earlier window, test on the next unseen window, then roll forward. TradeWell requires this OOS pattern for research architectures before any promotion to live TW Quant Score rankings.
Is walk-forward the same as NDCG@30?: No. Walk-forward OOS is a research gate. Live rankers are evaluated on NDCG@30 (NDCG@20 when the evaluated list is shorter than 30).
Does clearing walk-forward auto-promote a model?: No. Promotion is operator-only. A research name is not live because it appears in a paper, a Slack card, or a challenger job.
Is TradeWell investment advice?: No. TradeWell publishes research scores and forecasts, not investment advice, and does not tell users what to buy or sell.

Open the web app Get the iOS app