How we predict stocks

No. We deliberately don't use large language models for the prediction itself.

LLMs hallucinate confidently, are non-deterministic (same input → different output), are expensive at scale, and have no theoretical basis for predicting price movement.

Instead we use XGBoost — a gradient-boosted decision tree ensemble — on roughly 30 technical indicators. The output is deterministic: same input always yields the same prediction. We can also show you which features drove each call (see top_features in every prediction card).

All features are computed from past-only OHLCV data. There is no look-ahead leakage.

• Price-based: returns over multiple lookbacks (1d, 5d, 20d), z-scores, distance from moving averages • Volume: relative volume vs N-day median, on-balance volume • Volatility: ATR, realized volatility, Bollinger Band position • Momentum: RSI, MACD, rate-of-change • Microstructure: range/body ratios, gap features, intraday high-low spread • Calendar: day-of-week, day-of-month seasonality

The exact set is in backend/agents/statistical.py. We do not use fundamental data, analyst targets, or news scores in the prediction itself.

Walk-forward TimeSeriesSplit cross-validation, not random K-fold. Random K-fold would shuffle the time order and inadvertently let the model train on future data — the #1 silent mistake in financial ML.

Specifically: we train on the oldest N% of data, predict the next chunk, then expand the training window forward. Every prediction is evaluated against data the model has never seen.

The directional accuracy (DirAcc) we surface in every prediction card is this out-of-sample number — not a curve-fit training-set score.

Live track record at /predictions — every call is logged with its expiry, then verified against the actual close once the horizon passes.

Realistic expectations: on our verified log, daily directional accuracy has run around 50-55% on liquid US large-caps — our deepest sample. It is lower, sometimes below a coin-flip, on several smaller and Asian markets and at longer horizons. We don't claim 80%+ accuracy — anyone claiming that is either overfit, looking at noise, or lying.

Even a 55% directional edge is genuinely useful when compounded over many independent bets — but it is not a money printer, and it is not uniform across markets. If we wanted to mislead you we'd show a single cherry-picked ticker. Instead we publish every prediction we've ever made — wins, losses, and the full per-market breakdown — at /predictions.

Works best so far: US large-caps (deep history, liquid) on daily horizons with stable volatility regimes — our verified US directional accuracy leads every other market. Weaker so far on several Asian markets, where verified accuracy has trailed coin-flip; we surface the full per-market breakdown at /predictions rather than bury it.

Fails predictably on: • Earnings-day gaps — technical features can't anticipate fundamentals • Halted or suspended stocks — no recent data to train on • First-day IPOs — no history at all • Black swan events (COVID-style shocks, war, regulatory shutdowns) • Very thin / illiquid micro-caps — noise dominates signal

Each prediction is tagged low / medium / high confidence. A low-confidence call is explicitly not a trade signal.

Three reasons.

First, tabular financial features are XGBoost's home turf. The deep-learning vs gradient-boosting gap on small tabular data is well-documented: typically 1-2% in either direction, with XGBoost winning more often than not.

Second, Render's free-tier 512MB RAM ceiling won't fit TensorFlow. We could pay for more, but our pricing claim is that you can self-host this without burning $500/month.

Third, determinism and auditability. Tree models let us show you which features drove a specific prediction. Neural networks are harder to explain in a way a user can verify.

We use a deterministic lexicon scorer based on the Loughran-McDonald financial dictionary, plus our own multi-word phrase list. Each headline is scored ±1 and aggregated.

We explicitly do not use LLM-based sentiment. LLMs give different scores for the same headline on different calls, occasionally mark obviously negative news as positive when wrapped in ambiguous language, and are slow and expensive at scale.

A lexicon-based score isn't perfect — but it's reproducible, and we can show you exactly which words triggered the score. News sentiment is a side-signal in the UI; it does not feed into the price prediction itself.

Every prediction is a freshly-trained model. On each /predict call we:

1. Pull the latest OHLCV for the ticker 2. Compute features 3. Train a fresh XGBoost with walk-forward CV 4. Predict the requested horizon

This means the model always sees the latest data. No stale weights from a deployment last month. The cost is that predictions are slower than serving a pre-trained model — but for the trade-off in freshness, we accept the latency.

A point estimate: "AAPL will close at $X in 7 days," with explicit confidence (low / medium / high) and a Bullish / Neutral / Bearish directional read.

The point estimate is rounded to the nearest cent — but the confidence interval is much wider than that. Treat the number as a directional hint with explicit uncertainty, not as a target price. The action label and confidence band are more informative than the exact dollar value.

Don't take our word for it. Audit us.

Every prediction we have ever served is logged at /predictions, with the actual close filled in once the horizon passes. Per-ticker win rates are at /leaderboard. If our headline numbers don't add up, you will see it.

We can't hide stats that are computed from public log entries. That's the whole point.