Excerpt
Technical Analysis of Stocks & Commodities, Volume 44, Issue 10 (October 2026), "AI Agents" section. Read the full article in the magazine.
The question every strategy developer asks
Every strategy developer eventually asks: how did it perform in the future? A backtest looks good on the data used to build it. The real test starts the day after publication, when every new observation is genuinely new.
I gave an AI agent my own September 2017 article, "Profiting From Monthly Trading Cycles", and asked it to read it, rebuild the rules, freeze the parameters and test everything that happened after publication, from October 2017 to July 2026. Costs of 10 basis points per trade, nothing re-optimized.
What it found
The late-month effect survived: 105 trades on SPY, a 64.8% win rate and +0.49% per trade, and it was positive on DIA, QQQ and IWM too. The RSI, Bollinger and breakout filters added little or nothing consistently. Buy and hold still won on absolute return (242.6% against 59.2%), but the cycle was in the market only about 37% of the time, with a much shallower drawdown. Its practical value is as a timing filter, not as a replacement for market exposure.
The calendar effect survived. The technical indicators layered on top were inconsistent.
What it did not do
The agent did not take responsibility. A rolling high that includes today's bar, a calendar day confused with a trading session, a confident assumption the article never made: the researcher still has to catch these. The agent reads, codes and documents; the human defines the hypothesis, challenges the results and owns the conclusions.
Next in the series
Next: "EasyLanguage Meets the AI Agent" (Coming Soon) brings the agent's workflow into EasyLanguage and TradeStation and looks at how trading platforms are evolving.
Back to Publications Book a free call