tickle Implementation Timeline June 2026

Phases

Complete implementation timeline from baseline to profitable strategy

Phase Timeline

Summary

Phase 1: Technical Indicators

Status: Complete · Impact: +6.7% win rate

What Was Added

• 15 technical indicators (RSI, MACD, Bollinger, ATR, ADX, ROC, volume)
• 20 total features (5 price + 15 indicators)
• Dataset with feature normalization

Results

Key Finding: Technical indicators increased win rate by 6.7% and improved NLL by 27x. The model became more selective (95% fewer trades, better quality).

Files Created

• src/features.py — Technical indicators
• src/log_return_data_features.py — 20-feature dataset

Phase 2: Multi-Stock Training

Status: Complete · Impact: 46.7% win rate on META

What Was Added

• Training on AAPL, TSLA, MSFT simultaneously
• Per-stock normalization
• Generalization testing

Results

Key Finding: Multi-stock training achieved better NLL (-4.03) and higher win rate on test stock META (46.7%).

Phase 3: Architecture Experiments

Status: Complete · Impact: Conv1D best for this task

What Was Tested

• Transformer architecture
• LSTM encoder variants
• GRU predictor (performed worse)
• JEPA vs baseline comparison

Results

Key Finding: Transformer achieved best NLL, but Conv1D was selected for subsequent phases due to speed and similar performance in full configuration.

Phase 4: Longer Lookback

Status: Complete · Impact: Best NLL overall (-4.04)

What Was Changed

• Increased context window from 10 to 30 ticks
• Multi-stock training maintained

Results

NLL = -4.04 — best result across all phases. The longer context captured more temporal patterns without overfitting.

Phase 5: HMM Regime Detection

Status: Complete · Impact: 3-state regime models

What Was Added

• 3-state HMM for regime detection (ranging, up, down)
• Regime-specific models
• 23 features (20 base + 3 regime indicators)

Regime Distribution (AAPL)

Results

All 3 regime models achieved NLL ≈ -4.02. The ranging regime (82% of time) was the most common — skipping it could reduce trades significantly.

Files Created

• src/regime_detector.py — HMM-based regime detection
• src/log_return_data_regime.py — Regime-aware dataset

Phase 6: Combined Best Configuration

Status: Complete · Impact: Target achieved (40-47% win rate)

Configuration

• 30-tick context window (from Phase 4)
• Technical indicators (20 features from Phase 1)
• Multi-stock training (AAPL, MSFT, TSLA)
• Conv1D encoder (best performer)

Results

Target achieved: 40-45% win rate on both test stocks! Returns still negative, requiring Phase 7 optimization.

Files Created

• scripts/train_best_config.py — Combined best configuration

Phase 7: Performance Improvements

Status: Partially Complete

Phase 7A: Position Sizing

Status: Complete · Impact: +1.32% return on AMZN

Testing revealed the winning Kelly configuration:

kelly_fraction = 0.7
min_expected_return = 0.005

Key Finding: High min_expected_return filters low-quality trades. This configuration only works on AMZN.

Phase 7B: Regime-Aware Trading

Status: Complete — +0.58 Sharpe improvement on META-TSLA

Implemented spread-based regime detection on pair z-scores (not individual stocks). Uses Hurst exponent + rolling trend to classify regimes: Ranging (0), Trending Up (1), Trending Down (2), High Volatility (3).

Key insight: Regime filtering improved Sharpe on all 3 profitable pairs. META-TSLA achieved target (+0.5 Sharpe). Selective filtering skipped 2–4% of entry attempts during unfavorable regimes. Use --use_regime_filter flag to enable.

Phase 7C: Risk Management

Status: Not Started — Stop-losses and take-profits could improve further

Phase 7D: Ensemble Inference

Status: Complete — No improvement (models too correlated)

Phase 7E: Extended Testing

Status: Complete — Winning config only works on AMZN

Phase 8: Pairs Trading

Status: Running · Currently training on ada

What Was Added

• Multi-timescale correlation features (5, 10, 15, 30, 60, 120, 240 ticks)
• Pair-aware dataset with spread targets
• Modified JEPA for pair prediction

Pairs

AAPL-AMZN, AAPL-META, AAPL-MSFT, AAPL-TSLA,
AMZN-META, AMZN-MSFT, AMZN-TSLA,
META-MSFT, META-TSLA,
MSFT-TSLA

Status

Files Created

• src/correlation_features.py — Multi-timescale correlation computation
• src/log_return_data_pairs.py — Pair-aware dataset
• src/models/pair_models.py — Pair encoder, decoder, JEPA model
• scripts/train_pairs_model.py — Training script
• scripts/backtest_pairs_model.py — Backtesting script

Phase 9: Normalization Fix + Exit Sweep

Status: Complete · Impact: META-MSFT Sharpe 5.72, META-TSLA Sharpe 6.51

What Was Fixed

• AAPL normalization bug: Replaced global mean/std with per-pair normalization
• Exit timing: Grid-searched exit_zscore [0.1–1.5] across all 10 pairs
• GPU optimization: Scaled to 4.3M params (emb_dim 128→512), batch 8192, 8 workers

Final Results (Per-Pair Optimized)

Key finding: Per-pair normalization fixed the scaling issue but AAPL pairs remain unprofitable (21–28% win rate). This is now a learning problem, not a data problem — Phase 9C (per-pair adapters) addresses this.

Files Changed

• src/log_return_data_pairs.py — Per-pair normalization, shared memory tensors
• scripts/train_pairs_model.py — Scaled architecture (4.3M params)
• scripts/backtest_pairs_model.py — Per-pair exit parameter support
• scripts/sweep_exit_params.py — Exit z-score grid search

Phase 13: Online Learning of Trading Parameters

Status: Complete · Impact: Gradient adaptation +9% return, best risk-adjusted performance

What Was Implemented

Developed an online learning system for dynamic exit z-score thresholds using gradient-based parameter updates. The system adapts trading parameters based on real-time PnL feedback with regime-aware storage and meta-learning for adaptive learning rates.

Core Components

• OnlineParameterStore — Persistent per-pair, per-regime parameter storage
• GradientEstimator — Finite difference gradient computation on trade PnL
• ParameterOptimizer — Adam-style optimizer with gradient clipping and safety reverts
• SpreadRegimeDetector — Regime detection (ranging, trending, high volatility)
• MetaLearningOptimizer — Adaptive learning rates based on performance trends
• TransferLearningInitializer — Parameter initialization for new pairs

A/B Experiment Results (60 days, 3 profitable pairs)

*With 30% Kelly fraction (realistic position sizing). Higher Sharpe range with MIN_VOLATILITY_FLOOR=0.5.

Convergence Test Results

META_TSLA: ✓ CONVERGED
  Optimal: 1.50 | Final: 1.62 | Error: 0.12 (8% deviation)

META_MSFT: ✗ PARTIAL CONVERGENCE  
  Optimal: 1.50 | Final: 1.83 | Error: 0.33 (22% deviation)
  Issue: Gradient signal weak

AMZN_META: ✗ NO CONVERGENCE
  Optimal: 0.50 | Final: 2.30 | Error: 1.80 (360% deviation)
  Issue: Different optimal regime

Key Findings

1. Gradient adaptation works — B+ achieves best risk-adjusted returns (highest Sharpe, lowest drawdown)
2. Bandit explores more — C+ shows highest absolute returns but higher variance
3. Convergence varies — META_TSLA converged well, others need improved gradient estimator
4. Kelly fraction matters — 30% Kelly gives 90% growth with 33% variance vs 100% Kelly
5. Meta-learning needs time — Requires 50+ days (>10 parameter updates) to trigger

Files Created

• src/online_parameters.py — OnlineParameterStore, GradientEstimator, ParameterOptimizer
• src/meta_optimizer.py — MetaLearningOptimizer, TransferLearningInitializer
• scripts/validate_online_learning.py — Parameter convergence testing
• scripts/run_ab_experiment.py — A/B experiment framework

Key Learnings

1. Technical indicators are essential — +6.7% win rate improvement
2. 30-tick context > 10-tick — Captures longer patterns, better NLL
3. Multi-stock training helps — Generalizes better, NLL -4.03
4. Transformer ≈ Conv1D — Both achieve similar performance in full config
5. Regime detection works — All 3 regimes achieve NLL ≈ -4.02
6. Position sizing matters — Kelly=0.7, Min Return=0.005 achieved positive returns
7. Stock-specific patterns — Winning config only works on AMZN

Final Statistics