Prescia AI
Predict. Learn. Evolve.
Prescia is an autonomous, self-learning trading intelligence platform that combines institutional-grade machine learning with human-like market context awareness. Unlike traditional automated trading systems with fixed parameters, Prescia continuously learns from every trade outcome, adapts to changing market regimes, and autonomously optimizes its decision-making — without human intervention.
Prescia is built on a dual-brain architecture — two complementary engines that work in concert around the clock, each purpose-built for a different time horizon, both continuously evolving.
The Nightly Brain operates after the markets close, free from time pressure. It ingests end-of-day bars, fundamentals, macroeconomic indicators, news sentiment, and sector data from multiple institutional-grade sources. It engineers features across 40–60 dimensions per symbol, trains LightGBM models on 252 days of market history, generates high-conviction swing predictions across 1-day, 1-week, 2-week, and 4-week horizons across a universe of over 5,000 tickers, and runs a 26-phase pipeline that includes Monte Carlo simulation, Prediction Coroner autopsy, accuracy calibration, stress testing, brain update, and supervisor monitoring. By the time markets open, the system already knows what it thinks.
The Intraday Brain takes over during market hours. It tracks hundreds of symbols in real time, processing live 1-minute and 5-minute bars and completing a full intelligence loop — data fetch, feature engineering, model inference, policy decision, risk check, execution — in 2–8 seconds per cycle, with sub-100ms model inference per symbol. It runs 390 of these cycles every trading day. After close, it runs its own 14-phase post-market pipeline to consolidate what it learned before handing back to the Nightly Brain. Nothing is discarded. Everything feeds forward.
Prescia analyzes over 5,000+ tickers every night across a 26-phase pipeline. During market hours, hundreds of symbols are tracked in real time through 390 intraday cycles per day — each completing a full intelligence loop in 2–8 seconds with sub-100ms model inference. Every swing prediction is backed by 5,000 Monte Carlo simulation paths. The platform supports five trading horizons: 1-day, 1-week, 2-week, and 4-week swing positions, plus intraday execution.
5,000+
Tickers Analyzed
26
Pipeline Phases
5
Trading Horizons
390
Cycles / Day
5,000
Monte Carlo Paths
Every night, data flows in from Alpaca, Polygon, Alpha Vantage, Finnhub, and FRED. The 26-phase pipeline moves in a deliberate sequence: data ingestion → feature engineering → model training → prediction generation → Monte Carlo simulation → Prediction Coroner autopsy → accuracy calibration → portfolio stress testing → brain update → policy engine → bot rebalancing → insight generation → supervisor monitoring. Each phase depends on the last. No shortcuts.
The system maintains three feedback loop speeds: a fast loop that adjusts intraday parameters in real time, a medium loop that tunes confidence thresholds weekly, and a slow loop that triggers full model retraining monthly — or earlier when performance degrades below preset thresholds. The result is a platform that adapts continuously rather than waiting for a scheduled maintenance window.
The system never sleeps.
Every minute during market hours, Prescia completes a full intelligence cycle: live data is fetched, features are engineered on the fly, models run inference, policy decisions are generated, risk checks are applied, and execution decisions are made — all within seconds. Each of the 390 daily cycles feeds back into the system, continuously sharpening its edge.
After the close, the Intraday Brain runs its own 14-phase post-market pipeline — consolidating the day's intraday learnings, updating calibration maps, reconciling positions, and preparing a clean handoff to the Nightly Brain. Nothing is lost between sessions.
Twelve proprietary capabilities built from first principles — not assembled from off-the-shelf components.
Most algorithmic systems are tuned once at deployment and drift as market conditions change. Prescia doesn't wait for a developer to intervene. After every trade settles, the system examines the outcome and quietly adjusts its own confidence thresholds, position sizing, and exit strategies — within bounded, auditable parameter ranges.
This isn't blind hill-climbing. Every parameter change requires a minimum number of supporting trades, is validated per market regime, and is automatically rolled back if post-adjustment performance deteriorates. The system can only move in directions the data supports.
The result is a platform that gets sharper over time without human intervention — one that knows its own edge and adjusts how hard it bets based on where that edge currently lives.
Most systems track P&L. Prescia tracks why P&L happened. Every trade is logged with comprehensive metadata: entry confidence, expected versus actual return, market regime at entry, hold duration, and reason for exit. This metadata becomes the raw material for a continuous learning process that runs in the background of every nightly pipeline.
The system learns actual win rates per confidence bucket — not assumed win rates from backtesting. It identifies which regimes produce reliable edges and which produce noise. It optimizes exit timing based on real observed decay curves, not theoretical holding periods.
Over time, outcome attribution transforms the system's intuition. It develops a memory of what actually worked, in what conditions, and bets accordingly.
When a prediction is wrong, most systems simply record the loss and move on. Prescia performs an autopsy. The Prediction Coroner is a causal decomposition engine that runs on every resolved prediction and classifies the failure: was it noise, a model error, an underweighted feature, a regime shift, or an entirely unmodeled event? Each cause-of-death maps to a prescribed corrective action.
This distinction matters enormously. A loss caused by noise requires no response — it's expected variance. A loss caused by a regime shift may warrant temporary suspension of a strategy. A loss caused by an underweighted feature is a signal to retrain with adjusted feature importance. The Coroner knows the difference.
The output feeds directly into the nightly learning cycle, ensuring that the system learns from its mistakes rather than simply repeating them in a different market day.
Every swing prediction exits the pipeline with a full probabilistic risk profile, not just a point estimate. Two independent Monte Carlo engines simulate 5,000 price paths per symbol using Geometric Brownian Motion, with optional Merton Jump-Diffusion to model sudden gap risk and fat-tailed innovations to capture the reality that markets aren't Gaussian.
The outputs — Value at Risk, Conditional VaR, profit probabilities, and maximum drawdown estimates — are attached to every prediction before it reaches the policy engine. Position sizing is calculated against these risk numbers, not against expected return alone.
This means Prescia never bets a position size derived from optimism. Every size is derived from a quantified distribution of outcomes, including the tail.
A strategy that works in a trending bull market will bleed in a choppy sideways tape. Most systems apply the same rules regardless of environment. Prescia classifies the market into one of five regimes — Bull, Bear, Choppy, Panic, and Stress — and routes decisions accordingly.
Each regime triggers a different set of parameters: position sizing, strategy selection, stop-loss levels, and exposure limits are all regime-conditional. Transitions between regimes require multiple confirmations across independent signals to prevent whipsaw — the system doesn't react to noise, only to sustained shifts in character.
The regime layer is one of the most important filters in the pipeline. It is the difference between a system that performs well in the conditions it was tested on and a system that adapts to the conditions it actually faces.
No single model is right all the time, and no model's relative accuracy stays constant. Prescia runs multiple LightGBM models whose outputs are combined into an ensemble — but the weights of that ensemble are not fixed at training time. They adapt continuously based on each model's recent accuracy.
Seven-day and 30-day performance windows are tracked with exponential decay, so recent performance counts for more. When a model that was strong last month starts underperforming this week, its influence on the ensemble decreases automatically. When a previously underweighted model finds its footing in the current regime, it earns more weight.
The ensemble is always pointing at what's working now, not what worked at the last training run.
Prescia doesn't only learn from trades it took. It learns from trades it should have taken. High-confidence signals that were blocked — by capacity constraints, risk limits, or conservative parameter settings — are logged as missed opportunities. The system then calculates hypothetical P&L and tracks them forward.
If missed opportunities consistently outperform executed trades, the system surfaces this pattern and suggests parameter adjustments to capture more of that edge in the future. It is a form of counterfactual learning: examining the road not taken to understand whether the map needs updating.
For a system trying to maximize long-run performance, ignoring forgone gains is as much an error as ignoring realized losses.
Every adaptive system faces the same fundamental problem: how much time do you spend exploiting what you know works versus exploring whether something better exists? Commit too hard to exploitation and you get trapped in local optima. Explore too aggressively and you burn capital on experiments.
Prescia uses Thompson Sampling with Beta priors — a principled Bayesian approach to this tradeoff — across strategy selection, symbol universe rotation, parameter exploration, and model testing. The algorithm naturally concentrates allocation on high performers while maintaining a calibrated exploration budget.
The result is a system that keeps looking for better configurations without abandoning the ones that are currently working.
Markets operate across multiple timescales, and a platform that only adapts at one speed will always be out of sync with at least one of them. Prescia runs a three-speed feedback system designed to match the natural rhythm of market information: a fast loop that adjusts intraday parameters in real time, a medium loop that tunes confidence thresholds and strategy weights on a weekly cadence, and a slow loop that triggers full model retraining monthly.
Each loop has its own triggers: the fast loop responds to intraday signal quality, the medium loop responds to weekly win rate drift, and the slow loop fires either on schedule or immediately when performance falls below defined thresholds. Learning is not a scheduled event. It is a continuous process.
Risk management in most algorithmic systems is a single stop-loss. Prescia treats risk as a layered architecture: trade-level stops, daily loss limits, weekly and monthly drawdown caps, market condition gates that halt trading on VIX spikes, and a hard emergency stop with automatic triggers that fires if all other layers fail.
Each layer is independent, so a failure in one does not compromise the others. If the risk rails fail to load at startup, the system defaults to the most conservative settings — it cannot accidentally run without protection.
The design philosophy here is that drawdown control is not a constraint on returns — it is the precondition for surviving long enough to compound.
Prediction is not enough. A portfolio of individually strong signals can still destroy capital if those positions are correlated — all moving together in a tail event that no single-symbol model anticipated. Prescia runs scenario-based stress tests nightly: rate shocks, sector crashes, tail events, and correlation risk analysis across the full position universe.
The system doesn't just predict what will happen. It stress-tests what could go wrong. Those risk metrics feed directly back into position limits, preventing the portfolio from accumulating silent concentration risk that only reveals itself in a crisis.
This is the difference between a system that performs well under normal conditions and a system that is explicitly designed to survive abnormal ones.
Prescia does not assume its models are well-calibrated. It measures calibration continuously. The Accuracy Engine uses sliding-window analysis to track prediction accuracy across time, symbols, confidence levels, and market regimes — maintaining a live picture of where the system is sharp and where it is drifting.
These calibration maps feed directly into position sizing. When the system has been accurate in a given regime and confidence band, it bets larger. When calibration drifts — when the system is claiming high confidence but delivering low accuracy — it automatically pulls back. The size of every trade is proportional to verified edge, not assumed edge.
Most systems overbet when they are wrong and underbet when they are right, because they cannot tell the difference. The Accuracy Engine eliminates that blindness.
Prescia contains novel intellectual property protected by pending U.S. patent applications. The following inventions are covered:
Patent Status: Provisional application filed, full utility patent pending.
Experience the platform for yourself.
Launch Prescia AI →