Experiments

Experiment

Cicada

on pretend money, August 21 to October 2, 2026

A set of mechanical trading rules, run on pretend money and measured against the market every day for six weeks. Now retired.

Loading the numbers.

What this is
This is paper money. No real money is at risk, and none has been.pretend money
Not investment advice, not a fund, not an offer of anything. Nobody's money is being managed.not advice
The numbers come from the account itself and the day's closing prices, once a day, after the close.once a day

How it ended

We stopped the rules on October 2, 2026, six weeks in. Over those six weeks they lost 3.1% of the pretend account. The S&P 500 gained 0.8%, and just holding what the rules held gained 0.7%.

Afterwards we replayed the same six weeks against what the market actually did, to separate two questions: did the rules lose, or did we run them badly? The answer was both, in about equal parts. Run perfectly, the rules would still have lost about half as much, and still trailed just holding. The other half was the machinery: on a few mornings it never ran, some exits came minutes late, and the prices it got were worse than the plan assumed.

We also looked back over the past year. On paper the rules made money, but nearly all of it came from simply owning a few volatile funds through a strong market, which needs no rules at all. What the rules added on top was smaller than the trading costs we measured once they were running for real. And most of that year is history the rules were built from, so it could show that they fit the past, not that they work.

That left no reason to keep going: a small edge on paper, gone after costs, and a live record that agreed.

What could come next

Testing the idea on markets it has never seen: other funds, other exchanges, or older history than the data it was built from. Only a test like that can tell us something new.

Charging every backtest the real costs we measured live, before believing a single number it produces.

Making the machinery boring first. Before it is trusted with any rule, it should prove every morning that it ran, and say so plainly when it didn't.

The line above stops where the rules stopped, and it stays up.

What didn't work

Each study fixed its hypotheses in writing before it looked at any result, and ends with a list of what died. Every line below is drawn as far as its study got.

  1. Tick-by-tick patternsEverything faster than a minute died. 0 of 720 tests positive on the widest-spread names
  2. Machine learning on daily barsFound nothing the fixed rules had not already caught
  3. Relationships between stocksNothing survived both legs' costs
  4. Free outside dataAlready in the price. The delta was zero
  5. Prediction markets against stocksBoth directions were zero
  6. Settlement arithmetic on weather laddersExact, and worth $44 a day at the outside. Nobody on the other side of it
  7. Energy data against contractsNo-go at the first checkpoint, before the rest was run
  8. A forecasting model on weather laddersNo-go. Three of its five frozen numbers failed
  9. Prediction markets495 cells, the best of them breakeven. The forward test was abandoned at one night of thirty
  10. Weather ladders, replayedPassed its gate, then earned under a dollar a day in rehearsal against a $75 one
  11. A weather-ladder engineBuilt in shadow, parked before it ever reached its gate

Eight stopped at the bench. Three were watched without money behind them. None reached live trading.

These rulesAlready running when the studies began, and retired October 2, 2026. None of the studies beat them, and they did not beat just holding what they held.
Ask me about it