Machine learning in finance attracts as many fantasies as misunderstandings. This lesson aims at the opposite: explaining honestly what a model can do, and above all what it will never be able to do.
What a model actually does
A machine-learning model looks for statistical regularities between input variables (the “features”: momentum, volatility, valuation, volumes…) and an observed outcome (the return over N days). It has no understanding of the economy: it detects associations in past data.
Why finance is the worst playing field for ML
- The signal-to-noise ratio is catastrophic: almost all price variation is noise.
- The data is not stationary: market regimes change, and what worked yesterday stops working.
- The system is reflexive: if a strategy works and spreads, exploiting it destroys it.
- Data is scarce: only a few decades of history, where other fields have billions of examples.
The promise to run from
Any model announcing high accuracy in predicting prices is either overfitted or dishonest. An honest model shows a SMALL but statistically real edge — and that is already enormous, because it is that small edge, repeated, that builds performance.
Explainability: the real issue
A prediction without an explanation is unusable: there is no way to know whether the model has captured a real mechanism or an artefact of the data. SHAP values make it possible to measure each factor's contribution to a given prediction, and so to stay in control.
Earnnest's position
A model belongs in Earnnest only if it is measured, and it is withdrawn if it does not hold up. What is displayed remains a factual diagnosis: the app never says “buy”, because no model should, and because it stays neutral and non-personalised.
The right posture
A model is an assistant, not an oracle. It processes more data than you and without emotion; but it understands nothing of the world, is blind to what is not in its data, and it will be wrong. Use it to inform a decision, never to avoid making one.
À retenir
- ✓ML detects statistical regularities; it does not understand the economy.
- ✓Finance is hostile ground: massive noise, non-stationarity, reflexivity, little data.
- ✓Be wary of strong accuracy claims; demand explainability (SHAP).