tickle

a Joint-Embedding Predictive Architecture for stock price prediction

What is tickle?

tickle is a JEPA-based (Joint-Embedding Predictive Architecture) system for stock price prediction. Unlike generative models that predict exact prices, JEPA learns representations in a joint embedding space, making it more robust to market noise.

The system processes 1-minute OHLCV data with 30-tick context windows and 20 input features (5 price + 15 technical indicators). It predicts log-returns with uncertainty estimates, enabling position sizing based on expected risk-adjusted returns.

Key Results

stock	dataset	return	win rate	sharpe	trades
AMZN	test	+1.32%	43.9%	0.52	312
META	test	-2.17%	48.1%	-0.09	81
AAPL	train	-2.43%	39.7%	-1.27	224
TSLA	train	-5.82%	40.2%	-0.99	764
MSFT	train	-9.42%	37.5%	-1.33	24

The winning configuration (Kelly=0.7, Min Return=0.005) achieved +1.32% return on AMZN with 43.9% win rate. Notably, this configuration only works on AMZN — suggesting the model learned stock-specific patterns rather than generalizable features.

Architecture

Model Components

Encoder: Conv1D (fastest), LSTM (best MSE), or Transformer (best NLL)
Predictor: MLP with 128→128 hidden layers
Decoder: Log-return decoder predicting mean and variance
Loss: Negative Log-Likelihood (NLL) with variance weighting

Input Features

5 price features: Open, High, Low, Close, VWAP
15 technical indicators: RSI, MACD, Bollinger Bands, ATR, Volume EMA, etc.
All features normalized per stock

# Model architecture in PyTorch
context_encoder = Conv1DEncoder(
    input_dim=20, 
    hidden_dims=[64, 128], 
    embedding_dim=128
)

predictor = MLPPredictor(
    embedding_dim=128, 
    hidden_dims=[128, 128]
)

decoder = LogReturnDecoder(
    embedding_dim=128, 
    hidden_dim=64, 
    output_features=5
)

Trading Strategy

The trading strategy uses a Kelly-inspired position sizing approach with configurable parameters:

# Winning configuration
kelly_fraction = 0.7          # More aggressive than half-Kelly
min_expected_return = 0.005  # Filter low-quality signals
max_position = 1.0            # Full position cap

# Position calculation
raw_kelly = expected_return / variance
position_size = raw_kelly * kelly_fraction

Key Insight

The high min_expected_return threshold (0.005) filters out low-quality trades, while the aggressive kelly_fraction (0.7) capitalizes on the edge when signals pass the filter. This combination achieved positive returns where the baseline failed.

Implementation Phases

phase	description	status	result
Phase 1	Technical Indicators	complete	+6.7% win rate
Phase 2	Multi-Stock Training	complete	46.7% win rate
Phase 3	Architecture Experiments	complete	Conv1D best
Phase 4	Transformer	complete	overfit
Phase 5	HMM Regime Detection	complete	3 regimes
Phase 6	Best Config	complete	40-46% win
Phase 7A	Position Sizing	complete	+1.32% AMZN

Quick Start

# Activate environment
source venv/bin/activate

# Run training with best config
python scripts/train_best_config.py --symbol AMZN --device cuda

# Run backtest with optimized parameters
python scripts/train_and_backtest_fast.py \
    --symbol AMZN \
    --kelly-fraction 0.7 \
    --min-expected-return 0.005 \
    --use-features

Project Structure

tickle/
├── src/
│   ├── models/           # Encoders, Predictors
│   ├── features.py       # Technical indicators
│   ├── strategies.py     # Trading strategies
│   └── simulator.py      # Backtesting engine
├── scripts/
│   ├── train_best_config.py
│   ├── optimize_position_sizing.py
│   └── test_winning_config_all_stocks.py
├── stock_data/         # 1-minute OHLCV data
└── results/            # Experiment results