Trading Strategies

Multiple built-in strategies covering different market conditions and detection types:

Detection System

The platform uses a dual detection approach:

• Fast Mover Detection — Monitors real-time price action for rapid percentage moves, triggering strategy evaluation on qualifying symbols
• Slow Mover Detection — Identifies volume regime changes in quieter stocks that may signal emerging momentum

Signal Arbitration

When multiple strategies fire on the same symbol simultaneously:

• Priority-based signal arbiter resolves conflicts
• Cross-strategy correlation analytics prevent redundant overlapping trades
• Each signal includes confidence score and strategy metadata
• Per-strategy risk parameters control position sizing and exposure
• Detection-type filtering ensures strategies only fire on matching stock behavior (fast, slow, or both)

Indicator Engine

A declarative indicator subsystem shared across strategies and rule mining — indicators are computed once and reused, not recomputed per caller.

• Declarative registry — Indicators are registered with explicit dependencies; a topological dependency graph resolves compute order automatically
• Incremental append engine — New bars update indicator state in constant time per bar (O(1)), without full recomputation of historical series
• Demand-based lazy evaluation — Only indicators actually needed by the active strategy or mining run are materialized, keeping memory and CPU footprint lean
• Shared cache — Live-trading strategies and rule-mining searches read from the same indicator cache, preventing duplicate computation across paths

Strategy Fingerprinting

Before the rule-mining engine proposes new rules, the system first profiles what each existing strategy catches across historical data, identifies where coverage is thin, and generates curated seed candidates targeting those gaps.

• Historical replay profiling — Each strategy is replayed across historical data to capture what it catches and where
• Per-strategy coverage — Measures which market conditions each strategy detects, making its footprint explicit
• Cross-strategy overlap analysis — Surfaces where strategies redundantly cover the same opportunities versus where genuine gaps remain
• Seed-candidate generation — Turns coverage gaps into curated seed candidates that feed the rule-mining engine, directing discovery toward under-covered market conditions
• Dashboard integration — Fingerprinting runs appear in a dedicated dashboard with a multi-tab run-detail view (overview, coverage, cross-strategy, seeds)

The architecture diagram shows the live runtime data path; fingerprinting operates on historical replay and sits outside that path, so its absence from the diagram is by design.

Rule Mining Engine

Automated strategy discovery combining ML-seeded candidate generation, multi-phase beam search, and exit-quality scoring:

Approach

The engine builds trading rules through a phased search process, with an optional ML pre-seeding phase upstream:

0. ML Seeding (Optional) — A LightGBM model trains on historical trade data, extracts decision paths as seed candidates, and pre-selects promising condition combinations for the beam search to refine. Seed candidates can also arrive from Strategy Fingerprinting, targeting coverage gaps the existing strategies leave behind
1. Condition Screening — Evaluates individual indicator conditions across multiple categories — price-based, temporal indicators, event flags, and derivative / trajectory indicators that capture how indicator values change over time rather than only instantaneous levels
2. Condition Combination — Progressively combines top-performing conditions into multi-condition rules, pruning at each stage
3. Exit Optimization — Grid search across risk parameters (stop loss, trailing stop, profit targets, max hold time) to find optimal exit configurations per rule
4. Temporal Pattern Discovery — Explores multi-phase sequential patterns where conditions must fire in a specific order with configurable timing gaps

Carry-mode simulation — Outside the numbered pipeline, rules can optionally be evaluated with multi-day carry-over of open positions, enabling discovery of patterns whose entry is intraday but whose exit spans session boundaries.

Validation

• Time-series train/test split with chronological integrity — preserves temporal ordering for walk-forward evaluation
• Random train/test split with reproducible seeding — enables i.i.d. validation for rules where temporal structure is less load-bearing; the seed makes results deterministically replayable
• Walk-forward cross-validation to ensure rules generalize beyond training data
• Trade-set deduplication to remove redundant rules that trigger on the same opportunities
• Suspicion scoring to penalize rules that appear too good to be true

Exit Quality Scoring

A composite scoring mode evaluates rules by exit behavior rather than aggregate P&L alone, weighing multiple dimensions simultaneously:

• P&L — Net profit and loss
• Drawdown — Worst intra-trade drawdown
• Hold efficiency — P&L relative to time in the trade
• Exit type quality — Whether the exit hit a favorable target versus a defensive stop
• Peak-to-exit giveback — How much of the best intra-trade price was surrendered before exit

A scoring-mode selector lets users choose between P&L-centric and exit-quality-centric evaluation during discovery.

Rule Lifecycle

Discovered rules follow a defined promotion path:

1. Discovery — Mining engine surfaces candidate rules
2. Validation — Walk-forward testing confirms generalization
3. Campaign Integration — Multi-day automated discovery runs continuously explore new patterns
4. Live Promotion — Validated rules with per-rule exit configurations are promoted to live trading
5. Performance Tracking — Live results are tracked with analytics breakdown by rule
6. Degradation Detection & Demotion — Health scoring compares backtest vs live metrics per rule; degradation triggers toast notifications across all dashboard pages; underperforming rules can be demoted from live trading

Dashboard Surfaces

The Strategy Discovery dashboard exposes mining configuration and results in-place:

• Split-mode selector with a ratio slider lets users choose between time-series and random train/test split and tune the train/test ratio per run
• Carry-mode toggle controls whether simulations carry positions across sessions
• Capital-utilization statistics surface how much simulated capital a rule deploys and how concurrently its trades overlap
• Worst-trade analytics highlight the bottom-quantile outcomes so users see downside shape alongside headline performance

Exit Engine

Centralized exit logic shared across all strategies and promoted rules:

• Stop Loss — Fixed percentage-based stop loss
• Trailing Stop — Dynamic stop that follows price movement
• Scaled Profit Targets — Multi-tier partial profit taking at configurable levels
• Time-Based Exits — Maximum hold duration enforcement
• Price Freshness Validation — Stale price guards prevent exit decisions on outdated market data; buy orders use fresh Redis tick prices with automatic fallback

Each strategy and promoted rule can define its own exit configuration, enabling fine-grained risk management without global fallback behavior.

Broker Interface

Abstract BaseBroker interface enables seamless switching via configuration.

Alpaca

Interactive Brokers

Backtest Engine

Distributed backtesting with parameter optimization:

• Sweeps parameter combinations across all strategies in parallel
• Dedicated backtest workers in sharded mode for distributed execution via Redis work queues, with a finalize queue handling asynchronous result aggregation separately from the API process
• Tick data replay from QuestDB enables high-fidelity backtesting against historical market conditions
• Results stored with per-strategy analytics
• Best-performing configurations can be applied to live trading
• Real-time progress streaming via WebSocket to the frontend dashboard

Campaign Backtesting

Multi-day strategy comparison infrastructure:

• Aggregates performance across arbitrary date ranges in a single operation
• Per-date data pre-caching for worker efficiency
• Cross-day aggregation surfaces strategy consistency and parameter stability
• Partial failure handling — campaigns continue if individual dates fail
• Supports multiple discovery modes for different market behavior types

Production Correctness

Production-grade operation depends on handling real-world failure modes that simple code paths and backtests alone hide:

• WebSocket Auto-Reconnect — The market-data stream reconnects with exponential backoff on transient failure, preserving subscription state across disconnects so live strategies resume without manual intervention
• Redis Connection Manager — Auto-reconnect, keepalive heartbeats, and periodic health pings keep the pub/sub and cache layer available under network wobble; payload compression reduces bandwidth on high-throughput channels
• Structural Buy-Sell Linking — Trade accounting follows the actual order relationship between buys and sells rather than FIFO-by-timestamp, so P&L attribution matches the order-book truth even when exits fill out of submission order
• Split-Adjusted Historical Bars — Backtests and live price references use split-adjusted bars from the broker, preventing spurious signals and pricing errors around corporate actions

Frontend Pages

Notification System

Live notification bell wired to backend WebSocket events with severity-tiered delivery:

• Hybrid read/dismiss model — opening the dropdown marks notifications as read; individual items can be dismissed
• Redis-backed persistence with automatic TTL expiration for transient event data
• Backfill on reconnect ensures no events are missed during connection interruptions
• WebSocket connection status indicator on the bell icon (connected/reconnecting)

Database Schema (Key Tables)

Trading

Backtesting

Rule Mining

Authentication & Security

• JWT-based authentication with access and refresh tokens
• Extended session support for persistent dashboard access
• Protected API routes requiring bearer tokens
• PDT (Pattern Day Trader) compliance tracking

CLI Tooling

Administrative command-line interface built with Typer and Rich for database management, service control, operational tasks, and strategy fingerprinting — analyzing historical runs and generating seed candidates that feed the rule-mining engine.

Code Quality & CI/CD

• Linting — Ruff for unified formatting and lint checks
• Type Checking — MyPy in strict mode across the entire backend
• Testing — pytest with async support, property-based testing (hypothesis), time mocking, and HTTP mocking
• CI Pipeline — GitHub Actions runs lint, typecheck, and full test suite on every push and pull request
• Dependency Management — Renovate for automated dependency updates with CI gating
• Type Generation — OpenAPI-driven TypeScript type generation with CI drift detection