What I build, and what the tests said.
These pages document my research tools and historical tests. Every study belongs to a specific version. The planned indicator catalogue is separate from these historical results.
An AI market agent that has to earn every trade.
It scans markets, picks only strategies that have passed their tests, and sends each idea through risk checks that run separately from the strategy. A strategy can't produce a signal until a versioned, tested release of it is marked eligible. I built it with Claude Code.
- Data
- Market scanner
- Eligible strategies
- Independent risk checks
- Execution, demo only
Aurora: a rule set that looked brilliant, until it didn't.
VerdictNot supportedAurora runs the same rules as an MT5 expert advisor and as a TradingView indicator. On the two years I tuned it on, the backtests made +41% and +38%. On the eight years before, the same rules lost 19.2%.
Indicators first, then an EA that has to agree with them.
I write a rule as a TradingView indicator first, so I can see every signal on the chart. Then I rebuild it as an MT5 expert advisor to backtest it with real costs, and check that the two agree trade by trade before I trust either. That workflow, not any one indicator, is what the second course teaches.
Why the series is called Out of Sample
My own trading bot made +38% in a one-year backtest.
Gold, the Nasdaq and the S&P 500, October 2025 to September 2026. 250 trades.
The year before: +41%.
September 2024 to September 2025, 191 trades. Two good years in a row. This is usually where the sales page starts.
Then I ran the exact same rules on the eight years before.
October 2016 to August 2024: −19.2% over 971 trades, with a 26% drawdown.
I'd picked its settings using 2024–26 data.
A backtest on the data you tuned it on will nearly always look good. The only fair test is on data it has never seen. That's what out of sample means, and why this series has that name.
+38.2%
Aurora v12: three separate backtests of the same rules, each starting from $100,000, with commission and overnight financing included. Each line restarts at 0% where its test begins. Hover the chart for any day's value. Backtests describe the past. Read the full autopsy
Show the numbers as a table
| Period | Sample | Result | Trades |
|---|---|---|---|
| Oct 2016 to Aug 2024 | Out of sample | −19.2% | 971 |
| Sep 2024 to Sep 2025 | In sample | +41.0% | 191 |
| Oct 2025 to Sep 2026 | In sample | +38.2% | 250 |
360 opening-range breakout variants against the pass mark
One idea. 360 versions. One pass mark, written first.
The opening-range breakout on five stock indices: 72 versions each, ten years of 1-minute prices, the spread paid on every trade. To pass, a version needed t of 2 or more.
Every version, plotted.
Each dot is one version's t-statistic. Further right is better. None of the 360 crossed the line.
The best: 1.66, on the DAX.
Close enough to tempt anyone. With 72 tries on one market, though, close isn't the right question.
Luck alone would have beaten it.
Try 72 versions of an idea with no edge at all, and the best one is expected to show t of about 2.92 by chance. The DAX's 1.66 isn't even close to that.
0 / 360
t-statistics of all 360 variants, computed from each variant's trades over the full test period, spread included. The luck line is the best t expected across 72 variants with no edge. Backtests describe the past. Read the full autopsy
From research to a usable tool.
The indicator catalogue will identify the exact release, supported conditions and limitations. The learning kit teaches a general workflow for your own ideas. Neither offer turns a historical result into evidence of future returns.