Optiver :: Architecture

Navigation

tools tasks rules principles architecture

Unified Agent Layout

                unified_agent/
                |-- data_client.py         # OptiverDataClient: data access layer
                |-- tool_registry.py       # ToolRegistry: 84 tools organized by category
                |-- execution_context.py   # ExecutionContext: shared state
                |-- plan_builder.py        # PlanBuilder: regex -> execution plan
                |-- agentic_engine.py      # State machine orchestrator
                |-- feature_engineering.py # V2 feature generation
                |-- model_pipeline.py      # LightGBM training with time-series CV
                |-- optuna_tuner.py        # Hyperparameter optimization
                |-- walk_forward.py        # Walk-forward backtesting
                |-- feature_drift.py       # Feature drift detection
                |-- target_drift.py        # Target/concept drift detection
                |-- data_quality.py        # Data quality monitoring
                |-- monitoring_alerts.py   # Alert routing
                |-- smart_tools.py         # Meta-tools (SmartAnalyze, SmartMonitor)
                `-- report_templates.py    # Structured report outputs
            

Execution Flow

CLI + REPL

cli/main.py provides interactive commands and one-shot queries.

Planner

PlanBuilder compiles regex patterns into ordered execution steps.

Engine

AgenticEngine executes plan steps and updates ExecutionContext.

Reports

report_templates generate analysis, model, and monitoring summaries.

Data Source

Trading at the Close data is loaded from ~/Python/Optiver/OptFeatureViz/train.csv.

~480k rows 200 stocks 481 dates 17 columns