Models
A model thatcannot reproduceitself is not amodel
Training is the easy half. The half that decides whether anything works is proving the model in production produces the same probabilities as the model in the notebook — and that a label you can predict is a label worth predicting.
What the contour enforces
Training is a job, and the model is an artifact
A run produces a stored model with a reference, not a file in someone’s home directory. A strategy names the reference, so the thing that scored in the backtest and the thing scoring live are the same object rather than two builds that were once equal.
Parity is checked, not assumed
A model must reproduce its own predictions through the inference path before it is usable. This is not ceremony: a single swapped column moves probabilities by enough to change decisions, and column order is exactly the thing that differs between a training frame and a serving frame.
Features are the leak surface
A whole-series transform in a feature pipeline is the most efficient way to build a model that has seen the future. The same truncation discipline that applies to a coded strategy applies here, and it matters more, because a leaky model scores beautifully instead of obviously wrong.
The label decides the ceiling
The closer a label sits to actual profit-and-loss, the harder it is to predict — we measured that directly on a replica of a published approach: as the label moved toward PnL, the model’s discrimination fell, and none of the three reached break-even profit factor even without fees. A model with a great score on an easy label is a model that has learned something you cannot trade.
The result, since a page like this usually omits it
+3.9%, and gate-failed
The first model this platform trained and traded returned +3.9% on its run and then failed the promotion gate on the criteria set before it started. That is the whole story, and it is written here because the alternative — a page about ML capability with no number on it — is the shape of every other page like this.
Also: a data pipeline that deleted 12% of every dataset
Five unstable feature columns, four of them normalised by the candle body, produced NaNs that removed 12.4% of every training set — and the rows they removed were disproportionately quiet markets. The candles were complete. The pipeline was the problem, and finding that required looking at what was dropped rather than at what remained.
Questions about ML models
What can I train?
Anything your runtime supports on the feature matrix you specify, as a sandboxed job with the model stored and referenced. Anything sized by the data has to be an artifact — standard output is capped and the overflow is discarded silently.
Can a model place orders?
Only through a strategy that a human promoted, under a risk policy, through the same gate as everything else. A model is a scoring function; it is never an execution path.
Is two years enough data to train on?
For a small model on a short horizon, sometimes. For anything with many parameters, no — and the discipline that matters most on thin data is the out-of-sample one, not the architecture.
Train something and check its parity
Free paper accounts run the whole contour, including the checks that fail things.