I tried my hand at making a SAC model that trades stocks
Full disclosure I’ve shared this in a few discord channels. I thought I should cast a wide net if I am to get some feedback.
Hey yall
wanted to share a project ive worked on for the past 5 months
It initially started as a way to make money on the side, but turned into a research project. Looking for some feedback both on the code and research methodology. I didn’t go to school for AI and I’ve never done any research before (I got help from Claude).
it’s a Soft Actor-Critic (SAC) reinforcement learning agent that trades US stocks using features derived from OHLCV (open/high/low/close/volume), VWAP, and transactions data, sourced from Massive (formerly Polygon.io). Early versions (1–4) operated on 4-5 years of minute-level data; later versions (5–7) moved to 22 years of daily bars. Versions 1-6 traded one stock and version 7 traded a five-ETF basket (equities, treasuries, gold, oil, and the dollar)
It did not beat buy and hold
It made money: +12% across three independently seeded runs on a held-out 2024–25 test window. The catch is that simply holding the same basket returned ~29% over that window, at the same risk. It captured under half the return of doing nothing, for an identical drawdown.
I should also mention I was only able to backtest on a bull regime.
here’s the code and full write up
