Commit Graph

3 Commits

Author SHA1 Message Date
Youssef Aitousarrah c792b17ab6 fix(discovery): fix three scanner hang/validation bugs found in ranker_debug.log
1. executor.shutdown(wait=True) still blocked after global timeout (critical)
   The previous fix added timeout= to as_completed() but used `with
   ThreadPoolExecutor() as executor`, whose __exit__ calls shutdown(wait=True).
   This meant the process still hung waiting for stuck threads (ml_signal) even
   after the TimeoutError was caught.  Fixed by creating the executor explicitly
   and calling shutdown(wait=False) in a finally block.

2. ml_signal hangs on every run — "Batch-downloading 592 tickers (1y)..." never
   completes. Root cause: a single yfinance request for 592 tickers × 1 year of
   daily OHLCV is a very large payload that regularly times out at the network
   layer. Fixed by:
   - Reducing default lookback from "1y" to "6mo" (halves download size)
   - Splitting downloads into 150-ticker chunks so a slow chunk doesn't kill
     the whole scan (partial results are still returned)

3. C (Citigroup) and other single-letter NYSE tickers rejected as invalid.
   validate_ticker_format used ^[A-Z]{2,5}$ requiring at least 2 letters.
   Real tickers like C, A, F, T, X, M are 1 letter. Fixed to ^[A-Z]{1,5}$.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-02-20 22:35:42 -08:00
Youssef Aitousarrah 43bdd6de11 feat: discovery pipeline enhancements with ML signal scanner
Major additions:
- ML win probability scanner: scans ticker universe using trained
  LightGBM/TabPFN model, surfaces candidates with P(WIN) above threshold
- 30-feature engineering pipeline (20 base + 10 interaction features)
  computed from OHLCV data via stockstats + pandas
- Triple-barrier labeling for training data generation
- Dataset builder and training script with calibration analysis
- Discovery enrichment: confluence scoring, short interest extraction,
  earnings estimates, options signal normalization, quant pre-score
- Configurable prompt logging (log_prompts_console flag)
- Enhanced ranker investment thesis (4-6 sentence reasoning)
- Typed DiscoveryConfig dataclass for all discovery settings
- Console price charts for visual ticker analysis

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-09 22:53:42 -08:00
Youssef Aitousarrah 369f8c444b feat: discovery system code quality improvements and concurrent execution
Implement comprehensive code quality improvements and performance optimizations
for the discovery pipeline based on code review findings.

## Key Improvements

### 1. Common Utilities (DRY Principle)
- Created `tradingagents/dataflows/discovery/common_utils.py`
- Extracted ticker parsing logic (eliminates 40+ lines of duplication)
- Centralized stopwords list (71 common non-ticker words)
- Added ReDoS protection (100KB text length limit)
- Provides `validate_candidate_structure()` for output validation

### 2. Scanner Output Validation
- Two-layer validation approach:
  - Registration-time: Check scanner class structure
  - Runtime: Validate each candidate dictionary
- Added `scan_with_validation()` wrapper in BaseScanner
- Validates required keys: ticker, source, context, priority
- Graceful error handling with structured logging

### 3. Configuration-Driven Design
- Moved magic numbers to `default_config.py`:
  - `ticker_universe`: Top 20 liquid options tickers
  - `min_volume`: 1000 (options flow threshold)
  - `min_transaction_value`: $25,000 (insider buying filter)
- Fixed hardcoded absolute paths to relative paths
- Improved portability across development environments

### 4. Concurrent Scanner Execution (37% Performance Gain)
- Implemented ThreadPoolExecutor for parallel scanner execution
- Configuration: `scanner_execution.concurrent`, `max_workers`, `timeout_seconds`
- Performance: 42s vs 67s (37% faster with 8 scanners)
- Thread-safe state management (each scanner gets copy)
- Per-scanner timeout with graceful degradation
- Error isolation (one failure doesn't stop others)

### 5. Error Handling Improvements
- Changed bare `except:` to `except Exception:` (avoid catching KeyboardInterrupt)
- Added structured logging with `exc_info=True` and extra fields
- Implemented graceful degradation throughout pipeline

## Files Changed

### Core Implementation
- `tradingagents/__init__.py` (NEW) - Package initialization
- `tradingagents/default_config.py` - Scanner execution config, magic numbers
- `tradingagents/graph/discovery_graph.py` - Concurrent execution logic
- `tradingagents/dataflows/discovery/common_utils.py` (NEW) - Shared utilities
- `tradingagents/dataflows/discovery/scanner_registry.py` - Validation wrapper
- `tradingagents/dataflows/discovery/scanners/*.py` - Use common utilities

### Testing & Documentation
- `tests/test_concurrent_scanners.py` (NEW) - Comprehensive test suite
- `verify_concurrent_execution.py` (NEW) - Performance verification
- `CONCURRENT_EXECUTION.md` (NEW) - Implementation documentation

## Test Results

All tests passing (exit code 0):
-  Concurrent execution: 42s, 66-69 candidates
-  Sequential fallback: 56-67s, 65-68 candidates
-  Timeout handling: Graceful degradation with 1s timeout
-  Error isolation: Individual failures don't cascade

## Performance Impact

- Scanner execution: 37% faster (42s vs 67s)
- Time saved: ~25 seconds per discovery run
- At scale: 4+ minutes saved daily in production
- Same candidate quality (65-69 tickers in both modes)

## Breaking Changes

None. Concurrent execution is opt-in via config flag.
Sequential mode remains available as fallback.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-02-05 23:27:01 -08:00