algorithmic trading / ai agents / backtesting
How My AI Trading Bot Decides What to Buy and Sell
Austin Starks walks through the three steps his AI agent takes to decide what to trade, using the rules of his own live Public Portfolio Challenge account rather than a hypothetical.
Step one is writing the strategy. A strategy here is a deterministic set of rules, not a prediction: it runs once a week and only when the VIX is under 35, it drops any of its 19 candidate stocks trading below its 100-day average or carrying negative 3-month momentum, and it buys at-the-money call options one to two years out at 6% of the book each. The video shows the rule the agent actually wrote, then plots four days that book really filled a Robinhood buy against the stock's own price line.
Step two is testing it, and this is where the video spends its evidence. The strategy went through a four-fold walk-forward certification over 2022 to 2026. Each fold trains on one slice of history, blacks out the following two weeks so nothing leaks across the seam, validates on data the model has not seen, and is then measured on data after that. The four out-of-sample folds returned +112.7%, +151.7%, +583.7% and -4.7%, with maximum drawdowns of 36.4%, 34.8%, 53.5% and 83.8%. The same configuration won all four, which is what the test is actually looking for. The losing fold is shown rather than averaged away.
Step three is deployment. A paper book runs fully autonomously: the condition triggers, the configured action executes, and new holdings are generated. Live trading defaults to semi-automated, queueing orders for human review and approval, which is what stands between an account and a bug in the engine, an outage, or one bad afternoon.
The walk-forward figures are backtest results on historical data, not a live track record, and the price chart shows movement in the underlying stock rather than the option position's own profit and loss.
Transcript
0:00My trading bot buys call options once a week, and only when the VIX is under 35.
0:04Here's exactly how it decided that, in three steps.
0:08Step one. It generates a strategy.
0:11A strategy is a deterministic set of rules for when to buy and when to sell.
0:15It is not a prediction, and it is not a vibe.
0:18It is a rule that fires the same way every single time.
0:21In my Public Portfolio Challenge, it takes 19 stocks.
0:24It throws out anything trading below its 100-day average, or with negative 3-month momentum.
0:30Then it ranks what's left by momentum, and buys at-the-money calls, one to two years out, 6% of the book each.
0:37Step two. It tests the strategy.
0:40It trains on one slice of history.
0:43Then it blacks out the next two weeks completely, so no information leaks across.
0:46Then it validates on data it has never seen, and tests again on data after that.
0:51Four separate folds, over four years.
0:54And it isn't looking for the best return.
0:56It's looking for whether the same configuration wins every single time.
1:00On the Challenge, three of the four folds made money.
1:03The fourth lost money, and drew down 84%.
1:07You need to see that fold before you risk a dollar.
1:09Step three. We deploy it.
1:12Either to a paper trading book, or live to the market.
1:15On paper it's fully autonomous.
1:17The condition triggers, the action that was configured executes, and it generates our new holdings.
1:22Live, the default is semi-automated, so it queues the orders up for us to review and approve.
1:27That's what stands between you and a bug in the engine, an outage, or one catastrophic afternoon.
1:33So in a sense, the AI isn't a bot.
1:35It's just automating the stack of a quant that works on Wall Street.
Join the conversation
Loading conversation…