Mechanism

Using AI to predict stocks — the desk mechanism

Community skepticism is fair. Models overfit, leak, and decay. Here is the actual pipeline this desk uses, including what it cannot do.

Flow diagram from define through architecture, pros, cons, and fit — Define → architecture → pros → cons → fit.

1. Define

Target: short-horizon US equity direction and a small expected percent move. Inputs: public daily OHLCV closes from Stooq. Non-goals: exact price prophecy, broker execution, forex, AI-equity ticker pages.

2. Architecture

Three votes on the same feature vector: momentum (returns + SMA stack), mean-reversion (RSI14 + 1d snapback), volatility regime (20d vol z-score choosing follow vs fade). Net score → direction; magnitude scaled from recent |5d| move; confidence from agreement, capped 48–64.

3. Pros

Transparent features, fails closed, no paywalled mystery model for the free desk, fast print for liquid names.

4. Cons

Thin edge, regime breaks, corporate actions/gaps, Stooq outages, and live degradation vs backtests.

5. Fit

Best as a second opinion beside your own process — paper trade first. Poor fit if you need guaranteed fills, options greeks, or multi-asset spam signals.