AI Brain
The AI Brain panel runs the Sezgin Quant Core — a Rust ML service (RandomForest over 17 features, walk-forward cross-validation, and a backtest engine) for forecasting individual symbols and the wider market. Distinct from AI chat (which answers natural-language questions about your panels).
Open it from the left sidebar (eye icon).
Beta / experimental — known issue
AI Brain is experimental and not currently dependable as a live tool. As deployed in production the underlying Sezgin service ships with an empty data volume and its forecast path runs with updates disabled (no_update=True), so it lacks the minimum history the model needs (≥254 bars). In this state the endpoints return HTTP 500 errors and no forecast is produced. The panel and its actions are documented below for reference, but treat any output as provisional until the service is properly seeded and enabled.
Layout
- Eye widget — desktop: a central eye with 14 orbital action buttons. Mobile: a small eye on top of a 2-column action grid.
- Symbol search — fuzzy match across the symbols the core has indexed.
- Horizon picker — choose forecast horizon in days, months, or years.
FULL API SWEEP— chains a multi-action run (fetch → train → predict → backtest) in one click.- KPI tiles — last run's signal, target, return, confidence.
- Backtest leaderboard — best-performing runs by ranked metric.
- Run history — chronological list of past actions and their results.
Actions
Buttons green-tick once the corresponding endpoint has returned an OK in the current session.
| Button | What it does |
|---|---|
| Status | Core health probe |
| Symbols | Lists symbols the core has indexed |
| Fetch Data | Pulls OHLCV for the selected symbol |
| Update | Incremental refresh of an already-fetched symbol |
| Train | Fits the per-symbol model |
| Predict | Generates the next-horizon forecast |
| Backtest | Replays the strategy on historical data |
| Report | Composite output of recent runs |
| Validate | Cross-checks model output against held-out bars |
| Hindcast | Re-projects past predictions vs. actuals |
| Train Mkt | Trains the broad-market model |
| Mkt Signal | Generates the broad-market forecast |
| Random | Fetches a random symbol — useful for exploration |
| Update All | Refreshes every previously fetched symbol |
Dependency: Predict, Train, Validate, Mkt Signal, and Train Mkt need fetched data first. The panel automatically chains the prerequisite fetch when missing.
Horizon
Pick days / months / years. The horizon scales:
- Training window
- The forecast point that Predict returns
- The look-ahead used in Backtest and Hindcast
KPI tiles
After a successful predict / backtest:
- Signal —
BUY(green),SELL(red), orHOLD(amber). - Target — projected price at horizon.
- Return — projected % move.
- Confidence — model's self-rated confidence (0–100%).
Backtest leaderboard
Sorts past backtest runs by their primary metric (cumulative return). Click a row to load that run's details back into the KPI tiles.
When to use AI Brain vs. AI chat
| AI Brain | AI chat |
|---|---|
| Quantitative — signal / target / confidence | Conversational — answers and summaries |
| Per-symbol or market-wide forecasts | Reads from your active panels and live data |
| Action-driven (train / predict / backtest) | Question-driven |
| Sezgin Rust ML service | LLM with Atlas context grounding |
Required services
The Sezgin Quant Core is a separate service:
SEZGIN_API_URL— defaults tohttp://localhost:8088.SEZGIN_API_CONTEXT— Docker build path (defaults to../services/sezgin-api).SEZGIN_FETCH_FROM— historical fetch start, e.g.2018-01-01.
The compose stack starts it automatically if the directory is present. Without it, AI Brain still loads but every action returns an error from the core's health probe. Note that even when the service is running, the current production deployment returns HTTP 500 on the forecast endpoints (empty data volume + updates disabled), so it is effectively non-functional until seeded — see the beta note at the top of this page.
Tips
- Use Random to discover symbols you haven't trained yet.
- Update All is cheap if a symbol set was fetched recently; otherwise prefer one-by-one to control cost.
- Hindcast is the fastest way to sanity-check a freshly trained model — if past predictions don't line up with what actually happened, your training window is probably wrong.