Question 1

Is the ML model "predicting price"?

Accepted Answer

No — and we're explicit about that because the ML-trading space is full of overpromises. The output is a BUY / HOLD / SELL probability distribution over the next ~5 bars, summarized as an accuracy-weighted vote across the models — HOLD is a first-class outcome, not a hedge. Over many trades the probabilities should calibrate; any single trade can go either way.

Question 2

What models are used and how often are they retrained?

Accepted Answer

Five trained models on crypto — Random Forest, Gradient Boosting, LightGBM, XGBoost, Logistic Regression — and six on stocks, adding a stacking meta-learner. There's no primary/fallback hierarchy: every model votes, weighted by its validation accuracy. Each symbol + timeframe combination has its own trained model (e.g. BTC_4h.pkl, AAPL_1d.pkl). Models retrain on a rolling window so they adapt to regime shifts.

Question 3

Are the models calibrated?

Accepted Answer

Yes. The model probabilities are isotonic-calibrated on a held-out validation split, so a score of 65% is meant to behave like 65% over enough trades. On stocks, a guard falls back to the uncalibrated base model when calibration would actually degrade the fit — and the SHAP feature contributions are there if you want to contextualize any single stock score.

Question 4

How is this different from "AI signal services" on YouTube?

Accepted Answer

Three ways. First, we don't publish a single win-rate — that's an oversimplification of probabilistic output. Second, ML is one of the 9 frameworks — its verdict feeds the Cross-Tool Consensus alongside the other frameworks, never the whole answer. Third, on stock analyses every probability is decomposed into SHAP feature contributions you can audit.

Question 5

Plans?

Accepted Answer

ML predictions are included on Stock Pro (stock models), Crypto Pro (crypto models), and the Bundle (both, with a 7-day trial). See the pricing page for current rates.