Live Trader Action Decoder

Live Trader Action Decoder & Confidence Threshold Fix

Problem

Three bugs prevented the live trader from executing trades after deploying v5.3.0 (8-action) models:

1. 8-action decoder missing: multi_tf_predictor.py:_get_rl_prediction() only handled n_actions == 7, falling through to legacy 3-action for 8-action models. BUY_50% (action=2) decoded as SELL (direction=-1). Actions 3-7 all decoded as HOLD.
2. Markov system init: timeframe_selector.py passed chains as a list instead of Dict[str, MarkovChain], crashing all multi-TF predictions.
3. Confidence thresholds unreachable: Base 0.55 + regime adjustments (up to 0.65) exceeded structural max of 8-action softmax (~0.55).

Root Cause Analysis

8-Action Decoder

The correct decoder (interpret_model_action()) existed in live_trader.py:654 but multi_tf_predictor.py had its own inline decoder that only handled 7-action and 3-action. When models upgraded from 7 to 8 actions in v4.0.0, the multi_tf_predictor decoder was not updated.

Detection pattern: If model has n_actions=8 and logs show dir=-1 when model should be buying, or permanent dir=0, suspect action decoder mismatch.

Confidence Thresholds

With 8 actions, the maximum softmax probability is structurally lower than with 3 actions (uniform = 0.125 vs 0.333). The practical max observed was ~0.56. A base threshold of 0.55 blocked >98% of signals.

Detection pattern: If logs show >95% of directional predictions rejected by confidence threshold, the threshold is likely calibrated for a different action space size.

Solution

Files Modified

1. alpaca_trading/prediction/multi_tf_predictor.py:367 — Added n_actions == 8 branch matching the action space:

   0=HOLD, 1=BUY_25%, 2=BUY_50%, 3=BUY_75%, 4=SELL_25%, 5=SELL_50%, 6=SELL_75%, 7=CLOSE
   

2. alpaca_trading/signals/timeframe_selector.py:119 — Changed list to dict with MarkovChain wrapping:

   chains={"volatility": MarkovChain(vol_chain), "trend": MarkovChain(trend_chain), "timeframe": MarkovChain(timeframe_chain)}
   

3. scripts/live_trader.py:1050 — Lowered base_threshold 0.55 → 0.40, regime adjustments reduced, extreme vol gate 0.70 → 0.45.

Final Parameters

# Confidence thresholds (v5.3.3)
base_threshold = 0.40
regime_adjustments = {
    'low': -0.02,      # 0.38
    'subdued': -0.01,  # 0.39
    'normal': 0.0,     # 0.40
    'elevated': +0.03, # 0.43
    'high': +0.05,     # 0.45
    'unknown': +0.02   # 0.42
}
extreme_vol_confidence = 0.45  # For annualized vol > 60%

Verification

• CNM immediately placed a buy order on first prediction cycle after fix (was blocked for 23 hours before)
• WST changed from permanent dir=0 to dir=1 (correct BUY signal)
• CLSK no longer crashes with 'list' object has no attribute 'items'
• 169 tests passed, 0 failures

Key Lesson

When adding new action spaces to the training environment, grep for ALL action decoders across the codebase. There were two independent decoders: interpret_model_action() in live_trader.py (correct) and inline code in multi_tf_predictor.py (broken). The correct one was never called by the live prediction pipeline.

Prevention: The interpret_model_action() function should be the single source of truth, imported wherever action decoding is needed (moved to a shared module, not duplicated).