Active · 2025—2026

Prediction Alpha

An unattended prediction-market trading daemon: seven registered strategies, two venues, calibrated ML sizing, and a React operator dashboard.

System

Prediction Alpha is a single long-lived Python daemon that scans prediction markets, scores opportunities, sizes positions, and places orders without supervision. Seven strategies register against a code-level registry — cross-venue arbitrage, copy-tracking of large on-chain positions, a sportsbook-odds price oracle, hourly crypto brackets, a new-market scanner, an ML scanner, and an entertainment predictor. Kalshi is the primary venue; Polymarket is integrated alongside it for arbitrage.

Constraint

Prediction markets move faster than manual monitoring can handle. Acting on them continuously means ingesting news and public data, classifying it into market families, sizing positions under a bankroll model, and holding the whole thing inside hard risk limits — every one of which has to run unattended.

Contribution

Built one asyncio daemon around a strategy registry, so a new strategy is a config class plus an implementation class rather than a fork of the loop. Signals route through 23 categories with per-category confidence thresholds and a local FinBERT pre-filter; XGBoost with probability calibration and a Kelly-criterion sizer turn a scored signal into an order. Seven SQLite stores hold positions, deduplication, statistics, snapshots, and configuration. Risk limits are enforced in code, and a React operator dashboard exposes positions, trade audit, and strategy performance. The first-party tree is 51 Python modules, grouped below.

Architecture

Input

Kalshi · Polymarket marketsNews · search trends · sportsbook odds

System

Strategy registry · seven strategies23-category router · FinBERT filterCalibrated XGBoost · Kelly sizer

Outcome

Orders on two venuesSQLite positions · trade auditReact operator dashboard

Verified

Seven strategies in a code-level registry

Two venues integrated at the order-placement layer

23 signal categories with per-category thresholds

28 pytest modules across execution, state, and validation

Capabilities

Trading strategies & registryArbitrage, whale copy-tracking, sportsbook oracle, crypto brackets, early-market and ML scanners, plus the registry that binds config to implementation.

Venue & data-source clientsSigned REST against Kalshi, the CLOB SDK for Polymarket, and news, search-trend, and price feeds behind a common client shape.

State & persistenceSQLite stores for positions, deduplication, statistics, snapshots, and configuration, each with its own store module.

Signal classification & alertingA 23-category router, a local FinBERT sentiment pass, noise suppression, and outbound trade alerts.

Feature engineering & cachesFeature assembly over historical prices, news volume, and search interest, with caches that keep the hot path off the network.

ML models & position sizingXGBoost with probability calibration, an ensemble layer, a per-strategy crypto model, and a Kelly-criterion sizer.

Configuration & schemaTyped daemon and ML configuration with schema converters, so limits and thresholds are declared rather than scattered.

Order executionSeparate execution paths for arbitrage legs, ML-scored entries, and whale-follow orders.

Backtesting & validationWalk-forward validation, a fill simulator, and a metrics module for evaluating a strategy before it trades.