majiayu000 Post-Training Model Validation Workflow
Experiment Overview
| Item | Details |
|---|---|
| Date | 2024-12-27 |
| Goal | Establish systematic workflow for validating trained models before deployment |
| Environment | Windows, Alpaca API, GPU-native PPO models |
| Status | Verified |
Context
After completing GPU training (e.g., on Colab), models need systematic validation before deployment:
- Gating Assessment - Does the model meet quality thresholds?
- Backtesting - How does it perform on recent data?
- Walk-Forward Validation - Is performance consistent across time periods?
- Deployment Decision - Paper trading, live, or retrain?
Verified Workflow
Step 1: Extract Training Archive
Training runs produce a zip file with models and summary:
# Extract to training_archives/
unzip Alpaca_trading_trained_YYYYMMDD_HHMMSS.zip -d training_archives/YYYYMMDD_HHMMSS_extract/
# Key files:
# - training_summary_YYYYMMDD_HHMMSS.json (metrics per symbol)
# - models/rl_symbols/*.pt (trained models)
Step 2: Model Gating Assessment
Apply v2.4.5 thresholds to classify each model:
| Classification | Fitness | PF | Consistency | MaxDD |
|---|---|---|---|---|
| APPROVED | >= 0.70 | >= 1.8 | >= 85% | <= 8% |
| REVIEW | >= 0.50 | >= 1.3 | >= 65% | <= 15% |
| DROP | < 0.50 | < 1.3 | < 65% | > 15% |
from alpaca_trading.training.gating import assess_model_quality
classification, flags, use_checkpoint, cp_idx = assess_model_quality(
final_fitness=metrics['fitness_score'][-1],
final_pf=metrics['profit_factor'][-1],
final_consistency=metrics['consistency'][-1],
final_max_dd=metrics['max_drawdown'][-1],
fitness_history=metrics['fitness_score'],
)
IMPORTANT: Training MaxDD is a PROXY metric (reward volatility), not actual equity drawdown. Old reward_scale=0.1 caused inflated values (35-80%). New reward_scale=0.001 produces realistic values (5-15%).
Step 3: Copy Models for Testing
Copy approved/review models to models/rl_symbols/:
cp training_archives/YYYYMMDD_HHMMSS_extract/Alpaca_trading/models/rl_symbols/SYMBOL_1Hour.pt \
models/rl_symbols/
Step 4: Simple Backtest (30 days)
Quick sanity check on recent data:
# Set Alpaca API keys (NOT yfinance for crypto)
export ALPACA_KEYS_FILE=API_key_100kPaper.txt
python scripts/run_backtest.py \
--model models/rl_symbols/SYMBOL_1Hour.pt \
--days 30 \
--capital 100000
Expected output:
- Total Return (%)
- Max Drawdown (%) - Should be much lower than training proxy MaxDD
- Win Rate (%)
- Profit Factor
Step 5: Walk-Forward Validation (Critical)
Tests out-of-sample performance across multiple time periods:
python scripts/run_backtest.py \
--model models/rl_symbols/SYMBOL_1Hour.pt \
--days 180 \
--capital 100000 \
--walk-forward 5
Interpretation:
| Metric | Good | Marginal | Poor |
|---|---|---|---|
| Positive Folds | >= 4/5 (80%) | 3/5 (60%) | <= 2/5 (40%) |
| Sharpe Range | < 1.0 std dev | 1-2 std dev | > 2 std dev |
| Return Range | All positive | Mixed | Mostly negative |
Example output from UNIUSD validation:
PER-FOLD ANALYSIS
Sharpe Range: -3.67 to 2.36
Sharpe Mean: -1.18 (+/- 2.17) # High variance = inconsistent
Positive Folds: 2/5 # Only 40% profitable
This indicates the model performs well in some market regimes but poorly in others.
Step 6: Deployment Decision
| Walk-Forward Result | Action |
|---|---|
| >= 4/5 positive folds, low variance | Deploy to LIVE |
| 3/5 positive folds, moderate variance | Deploy to PAPER for monitoring |
| <= 2/5 positive folds, high variance | RETRAIN with new parameters |
| Consistent losses | DROP model, investigate training data |
Failed Attempts (Critical)
| Attempt | Why it Failed | Lesson Learned |
|---|---|---|
| Using yfinance for crypto | yfinance doesn't support UNIUSD, BTCUSD etc | Always use Alpaca API for all symbols |
| Trusting training MaxDD | Old reward_scale=0.1 caused 35-80% phantom MaxDD | Backtest shows real MaxDD (2-5%) |
| Simple backtest only | Overlaps with training data, not out-of-sample | Walk-forward validation is essential |
| Deploying after gating only | Gating uses proxy metrics from training | Real validation requires backtesting |
Key Insights
Proxy vs Real MaxDD - Training MaxDD is from reward volatility, not equity. Real backtest MaxDD is typically 5-10x lower than training proxy.
Walk-forward is essential - A model can look good on aggregate metrics but fail in specific market regimes. Walk-forward reveals this.
Fold consistency matters - A model with 2/5 positive folds but high total return is being carried by one lucky period. Not reliable.
Alpaca API for all data - yfinance doesn't support crypto. Use
ALPACA_KEYS_FILEenvironment variable to specify API credentials.Time per fold - Each walk-forward fold takes ~7-8 minutes for 253 bars. 5-fold validation takes ~35-40 minutes total.
Commands Reference
# Quick backtest (30 days, recent data)
ALPACA_KEYS_FILE=API_key_100kPaper.txt python scripts/run_backtest.py \
--model models/rl_symbols/SYMBOL_1Hour.pt --days 30
# Walk-forward validation (180 days, 5 folds)
ALPACA_KEYS_FILE=API_key_100kPaper.txt python scripts/run_backtest.py \
--model models/rl_symbols/SYMBOL_1Hour.pt --days 180 --walk-forward 5
# Extended validation (365 days, 10 folds)
ALPACA_KEYS_FILE=API_key_100kPaper.txt python scripts/run_backtest.py \
--model models/rl_symbols/SYMBOL_1Hour.pt --days 365 --walk-forward 10
References
scripts/run_backtest.py: Backtest engine with walk-forward supportalpaca_trading/backtest/walk_forward.py: Walk-forward validation implementationalpaca_trading/training/gating.py: Model quality assessmentalpaca_trading/training/archive.py: Training archive management- Training archive:
training_archives/YYYYMMDD_HHMMSS_extract/