Prescia AI
From raw data to actionable intelligence.
Prescia operates on a dual-brain architecture — two purpose-built engines running continuously around the clock. The Nightly Brain is the swing intelligence engine: it runs a 26-phase overnight pipeline, engineers features, trains models, generates swing predictions, runs Monte Carlo risk simulations, and calibrates itself against every prior outcome. The Intraday Brain is a completely separate real-time engine that takes over during market hours, executing 390 intelligence cycles per trading day before handing off its learnings back to the Nightly Brain after close. Nothing is wasted. Everything feeds forward.
Prescia covers five distinct time horizons across its two engines — four swing horizons handled by the Nightly Brain, and intraday execution handled by the Intraday Brain.
Short-term overnight swing positions. Predictions generated end-of-day, capturing overnight momentum and next-session mean-reversion setups.
Handled by: Nightly Brain
Multi-day swing positions anchored to weekly momentum patterns, relative-strength signals, and short-cycle mean reversion.
Handled by: Nightly Brain
Medium-term positions designed to capture sector rotation, trend continuation, and broader risk-on/risk-off transitions.
Handled by: Nightly Brain
Longer-horizon positions based on fundamental shifts, macro indicator transitions, and sustained institutional flow patterns.
Handled by: Nightly Brain
Real-time execution during market hours. The Intraday Brain processes 1-min and 5-min bars, completing a full intelligence loop 390 times per trading day.
Handled by: Intraday Brain
Every night after markets close, the Nightly Brain runs a 26-phase pipeline. The phases are grouped into seven logical stages — each building on the last.
End-of-day OHLCV bars, fundamental ratios, macroeconomic indicators (FRED), news sentiment scores, and sector/index data are pulled from Alpaca, Polygon, Alpha Vantage, and Finnhub. All sources are normalized, validated, and stamped before entering the pipeline.
40–60 features are computed per symbol: price momentum, volatility regime, volume profile, cross-asset correlation, sector breadth, and calendar effects. Features are winsorized, z-scored, and stored in the feature cache for model consumption.
LightGBM models are trained on a rolling 252-day window for each horizon (1d, 1w, 2w, 4w). Ensemble inference runs across all models, with dynamic weighting that adjusts each model's contribution based on recent calibrated accuracy.
Swing predictions are generated for every symbol across all four horizons. Each prediction carries a direction, magnitude estimate, confidence score, and signal tier. Only signals clearing the calibrated confidence threshold are promoted to the active signal set.
Monte Carlo simulation runs 5,000 paths per symbol to estimate VaR and CVaR. Portfolio-level stress tests simulate correlated drawdown scenarios. Correlation risk analysis flags concentration in sectors, factors, and co-moving symbols.
The Prediction Coroner autopsies every closed prediction, classifying outcomes into noise, model error, regime shift, underweighted feature, or unmodeled event. The Accuracy Engine updates calibration maps across time, symbol, confidence tier, and market regime. The brain state is updated with the new calibration weights.
The policy engine translates predictions and risk scores into sizing decisions. Bots are rebalanced according to updated signals. Insight summaries are generated for dashboard display. The supervisor performs a final system health check before the pipeline marks complete.
The Intraday Brain is a separate real-time service. It operates entirely during market hours, running its own continuous intelligence loop independent of the swing pipeline.
Each of the 390 cycles per trading day executes the same deterministic sequence: real-time bar fetch → feature engineering → model inference → policy decision → risk check → execution. The entire loop completes in 2–8 seconds end-to-end, with sub-100ms model inference per symbol. The system processes 1-minute and 5-minute bars in parallel, maintaining separate feature vectors for each bar resolution.
2–8 seconds total · sub-100ms model inference
After the market closes, the Intraday Brain does not simply shut down. It runs its own 14-phase post-market pipeline to consolidate everything it learned during the session. This pipeline reconciles intraday positions, updates calibration maps based on the day's signal outcomes, recalibrates confidence thresholds, and packages the learnings for handoff. The Nightly Brain ingests this handoff at the start of its pipeline, ensuring that intraday intelligence flows into swing predictions — and vice versa.
Prescia does not have a single learning cycle. It operates on three feedback loops running at different speeds, each targeting a different type of performance degradation.
Continuously monitors intraday signal quality. If a model's signals degrade mid-session — measured by live P&L attribution vs. expected value — parameter adjustments are made within the same trading day without requiring full retraining.
Reviews rolling win rate drift, confidence calibration drift, and strategy-level Sharpe degradation. Confidence thresholds are re-tuned and ensemble weights are adjusted to favor models currently performing well relative to their historical baseline.
Full model retraining on the most recent 252-day rolling window. Triggers automatically when performance drops below a configurable degradation threshold — it does not wait for the monthly schedule. Feature importance rankings are reviewed and weak features are pruned.
Every closed prediction is autopsied. The Prediction Coroner assigns each outcome a cause-of-death classification: correct, noise, model error, regime shift, underweighted feature, or unmodeled event. Each classification maps to a prescribed action: retrain, feature exclusion, threshold adjustment, or regime weight update. Failures do not just get counted — they get explained and acted on.
The Accuracy Engine maintains calibration maps across four dimensions: time (rolling windows), symbol (individual ticker calibration), confidence tier (how well each confidence band predicts actual outcomes), and market regime (bull, bear, sideways, high-volatility). A prediction with 80% confidence should be right roughly 80% of the time in the conditions where it was made. The engine continuously verifies this and corrects drift before it compounds.
Risk is enforced through a five-layer independent rail system. Each layer operates independently — a failure in one layer does not compromise the others. If risk rails fail to load, the system defaults to its most conservative settings automatically.
Trade-Level Stops
Hard stop-loss and take-profit levels set per position at entry. These execute automatically and cannot be overridden by any higher-level signal.
Daily Loss Limits
If aggregate realized losses for the day exceed the configured daily loss cap, new position entries are halted for the remainder of the session.
Weekly / Monthly Drawdown Caps
Rolling drawdown caps at weekly and monthly windows. If portfolio drawdown exceeds the cap, the system reduces exposure aggressively and shifts to cash-preservation mode.
Market Condition Gates
VIX spikes, circuit breaker events, and extreme spread widening trigger trading halts automatically. The system detects these conditions in real time and suspends execution without human intervention.
Hard Emergency Stop
A top-level circuit breaker that forces liquidation and halts all activity if predefined catastrophic thresholds are crossed. Independent of all other layers — activates even if other risk rails have failed to load.
Fail-safe default: If risk configuration fails to load at startup, the system does not attempt to trade with undefined limits. It initializes with the most conservative pre-set profile and alerts the supervisor. All five layers must be confirmed active before live execution is permitted.
Everything described on this page runs live, continuously, and autonomously — every market day.
Launch Prescia AI →