v1.0.0 — Complete Ablation Study: DRL Trading with Sentiment Analysis

FinRL Deep Reinforcement Learning — v1.0.0

First stable release of the complete ablation study benchmarking DRL trading agents
across architectures, data sources, stock universes, and capital levels.

What's included

• 24 trained models across 2 universes (30-stock, 50-stock), 4 architectures,
  and 3 capital levels ($1M / $100k / $10k)
• utils.py — shared evaluation utilities for all notebooks
• Complete metrics — Sharpe ratio, total return, max drawdown, Calmar ratio,
  win rate, buy-and-hold comparison, market regime analysis
• GitHub Pages dashboard — interactive results at
  https://tyhobbs.github.io/FinRL_Deep_Reinforcement_Learning/

Key Results

Metric	Value
Best Test Sharpe	3.111 (30-Stock VGG Baseline $10k)
Models beating buy-and-hold	18 / 24
Best avg architecture	VGG + Alpaca (avg Sharpe 2.089)
Optimal capital level	$100k (avg Sharpe 2.035)

Key Finding

Contrary to the initial hypothesis, VGG-based architectures outperformed the
Cross-Stock Transformer (avg Sharpe 2.089 vs 1.568). Capital constraint level
is the dominant performance factor — $100k models achieve the highest average
risk-adjusted returns across all architectures.

Training Details

• Training period: 2020-01-01 → 2024-01-01
• Test period: 2024-01-01 → 2025-01-01
• Algorithm: PPO via Stable Baselines 3
• Sentiment: Polygon.io + FinBERT (full historical coverage)
• Transaction costs: 0.15% per trade (0.1% commission + 0.05% slippage)
• Risk-free rate: 5% annualised (3-month T-bill)

Architecture Summary

Architecture	Avg Test Sharpe	Best	Worst
VGG + Alpaca	2.089	2.726	1.570
VGG Baseline	1.939	3.111	1.001
VGG + FinBERT	1.892	2.350	1.273
Transformer	1.568	1.895	0.629

What's not included

Model weights (.zip files) and sentiment cache (.pkl files) are excluded
from this release due to file size. See README for instructions on
reproducing results from scratch.