Video summary
I Tested 26,834 Trading Strategies With AI. Here's What I Learned.
Main summary
Key takeaways
Summary
The presenter says he tested more than 26,000 trading-strategy variations—26,834, according to the title—on Nasdaq data and found only 29 that survived his tests. He argues that many strategies showcased online appear profitable over selected periods but fail when tested across longer histories or against chance.
Testing Process and Lessons
- The presenter says he trained strategies on historical data and tested them on years they had not seen. One test covered 16 years of Nasdaq data.
- Of 684 versions of common intraday strategies—including breakouts, moving averages, and indicator-based setups—the best reportedly performed no better than random.
- A YouTube strategy advertised as having a 70% win rate looked strong over the two months shown in the video, but was no better than a coin flip in the presenter’s 16-year test. Another strategy advertised as having 57% winners produced only 17%–39% wins when he reran its rules.
- A strategy that made money from 2024 to 2026 lost money in 12 of the 14 years from 2010 to 2023, according to the presenter. He cautions against judging a strategy based on only a few recent years.
- Testing many variations increases the chance of finding a seemingly successful result by luck. He recommends recording every tested setting and filter to reduce undocumented tweaking and overfitting.
- He says each strategy should beat 500 random strategies trading on the same day with the same stop-loss rules. 28 of his 29 winners passed; one did not and was removed from his portfolio.
- He recommends checking whether results are robust to small parameter changes. For example, if a strategy works only with a 40-session setting, but not 39 or 41, he treats that as a likely coincidence.
- Simulated trades should be checked to confirm they could actually have been filled. In one bot, comparing trades with one-minute prices showed that some assumed fills were impossible. Removing those trades cut reported profits by 74%, so he abandoned the bot.
- He also warns that a chance benchmark can be flawed. One of his luck tests effectively used future information and entered earlier than the real strategy. Comparisons should use identical rules, including the same stop size.
- Even a promising strategy may require 1.5 to 8 years of forward testing to establish an edge, he says. He paper-trades strategies while collecting forward-test data.
Strategies and Market Observations
These are the presenter’s reported Nasdaq backtest findings, not universal conclusions:
- A recurring pattern among his winners was an early-day breakout, a tight stop, and holding until the end of the day without a take-profit target.
- In his results, adding a take-profit, moving the stop to breakeven, or using a trailing stop worsened the tested statistics.
- Breakouts in the direction of the opening gap—when the market opens above the prior close—performed better in four of his winning strategies.
- After a major trend day, breakouts in the opposite direction were stronger in his tests; breakouts in the trend’s direction produced roughly zero.
- Most reversal approaches he tested—including fading breakouts, trends, or opens—failed. The exception he cites was fading large post-weekend gaps on Monday after the open.
- Continuation trades performed better when the day’s first move rose without a substantial pullback. He says the size of the pullback mattered more than how many minutes the initial move lasted.
- He says the same winning strategies returned close to zero when tested on the S&P, Dow, and gold, and that the results were specific to the Nasdaq.
Performance, Prop Firms, and Risk
- The presenter says his live bot loses about two trades out of three, but its average win is roughly three times its average loss. He reports that its backtest was profitable in each year from 2021 through 2026.
- He reports that a live-account bot made about $10,000 in 2025, versus $50,000 in previous years. He says several strategies that had been winning went flat starting in 2025, so he now requires new strategies to show their 2025 performance.
- In his simulated prop-firm evaluations, a near-zero-edge coin-flip strategy passed 27% of the time. He says his real strategies passed only somewhat more often, and that the funded stage—not the evaluation—is where he sees the real edge.
- He cautions that strategies with roughly 30% win rates and large winners may suit a regular account but can struggle with prop-firm trailing drawdowns and losing streaks. He says prop-firm accounts may favor strategies closer to a 50% win rate with smaller take-profits.
- His “Oracle” strategy reportedly trades about once a week. In forward testing beginning September 28 (year unspecified), it lost on three consecutive days, then made 4R on a short trade held to the close on the fourth day, more than offsetting those losses.
Main Takeaway and Disclosures
The presenter’s central recommendation is to test a strategy on unseen years and require it to beat a chance benchmark using exactly the same stop rules. He also emphasizes realistic execution checks and cautions that attractive results from short periods may reflect luck or overfitting.
He says he cannot code and built the tests by prompting Claude in plain English. He describes the findings as his personal data and says they may or may not be right; the results are not presented as proof of a general market edge. The spoken subtitles do not include an explicit “not financial advice” statement.
Assets and Instruments Mentioned
- Nasdaq (the principal market tested)
- S&P
- Dow
- Gold
- Prop-firm trading accounts and evaluations
No specific ticker symbols, ETFs, or cryptocurrencies are named.
Presenters and Sources
- Steven Ace (presenter; channel: Steven Ace Trades)
- Claude (AI assistant used to help build and run the tests)
- The presenter also discusses strategies and performance claims encountered on YouTube and Instagram.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.