Strategy

Kelly-inspired position sizing and backtesting methodology

Position Sizing Overview

The trading strategy uses a Kelly-inspired position sizing approach with configurable parameters. Unlike fixed-position strategies, this method adjusts exposure based on the model's confidence and predicted risk-adjusted returns.

Core Formula

# Kelly-inspired position sizing
raw_kelly = expected_return / variance
position_size = raw_kelly * kelly_fraction

# With constraints
if expected_return < min_expected_return:
    position_size = 0  # Skip trade

position_size = min(position_size, max_position)  # Cap exposure

Key Parameters

parameter	description	baseline	optimized
`kelly_fraction`	Fraction of Kelly to use	0.5	0.7
`min_expected_return`	Minimum expected return threshold	0.0005	0.005
`max_position`	Maximum position size	1.0	1.0

Position Sizing Parameter Sweep

Winning Configuration

Phase 7A testing discovered the configuration that achieved positive returns:

# Winning configuration (achieved +1.32% on AMZN)
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

# This configuration:
# - Uses 70% of Kelly (vs typical 50%)
# - Requires 0.5% minimum expected return (vs 0.05%)
# - Results in fewer but higher-quality trades

Why It Works

The high min_expected_return threshold (0.005 = 0.5%) filters out low-quality trades. While this reduces trade frequency, the surviving signals have higher expected edge. The aggressive kelly_fraction (0.7) capitalizes on this edge when signals pass the filter.

Key insight: Win rate improved from 35% to 44%, but more importantly, the quality of winning trades improved — leading to positive overall returns.

Performance Results

configuration	stock	return	win rate	sharpe	trades
Kelly=0.7, Min=0.005	AMZN	+1.32%	43.9%	0.52	312
Kelly=0.7, Min=0.005	META	-2.17%	48.1%	-0.09	81
Kelly=0.5, Min=0.0005	META	-3.09%	58.6%	-0.12	29
Baseline (unoptimized)	AAPL	-17.8%	28.5%	-56.21	3280

The winning configuration only works on AMZN — suggesting stock-specific pattern learning.

Backtesting Engine

The backtesting engine simulates trading with realistic constraints:

Key Features

Walk-forward testing: Train on earlier data, test on later data
Position sizing: Kelly-inspired with configurable parameters
Transaction costs: Configurable slippage and commission
Risk metrics: Sharpe ratio, max drawdown, win rate

Running Backtests

# 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 \
    --device cuda

# Test on all stocks
python scripts/test_winning_config_all_stocks.py \
    --kelly-fraction 0.7 \
    --min-expected-return 0.005

Performance Metrics

Financial Metrics

metric	formula	interpretation
`win_rate`	profitable_trades / total_trades	% of trades with positive P&L
`return`	final_value / initial_value - 1	Total percentage return
`sharpe`	mean(return) / std(return)	Risk-adjusted return (annualized)
`trades`	count(positions)	Number of positions taken

Prediction Metrics

metric	description	target
`dir_accuracy`	% correct direction predictions	> 50%
`dir_accuracy_1pct`	% correct for moves > 1%	> 50%
`nll`	Negative log-likelihood (lower is better)	Minimize
`mse_close`	Mean squared error on close price	Minimize

Risk Management

Current Approach

Position sizing via Kelly criterion (with fractional Kelly for safety)
Minimum expected return threshold filters low-quality signals
Maximum position cap prevents over-concentration

Future Improvements

Phase 7C (Risk Management) was planned but not completed. Potential enhancements:

Stop-losses: Automatic exit at predetermined loss levels
Take-profits: Lock in gains at target levels
Trailing stops: Protect profits as they accrue
Regime-aware sizing: Reduce size in high-volatility regimes

Stock-Specific Behavior

A critical finding from Phase 7E: the winning configuration only works on AMZN.

stock	dataset	return	win rate	notes
AMZN	Test	+1.32%	43.9%	Works!
META	Test	-2.17%	48.1%	Higher win rate, negative return
AAPL	Train	-2.43%	39.7%	Training data
TSLA	Train	-5.82%	40.2%	High volatility hurts
MSFT	Train	-9.42%	37.5%	Worst performance

Interpretation: The model learned patterns during training (AAPL, MSFT, TSLA) that generalize to AMZN's microstructure but not META's. This suggests the model captured stock-specific rather than universal patterns.

Quick Start

# 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

# Test on all stocks
python scripts/test_winning_config_all_stocks.py