Each line names the kind of idea and the reason it closed. We leave out parameters, stock lists and timing on purpose; the lesson is in the reason.
Each idea was judged against a bar: usually an index fund or a simpler method, after real costs and tax. Some were closed by a full test, frozen before we read its result. Others were closed at an earlier, exploratory look, because the evidence already fell short. Where a test could only have seen a large effect, we say so.
Costs and taxes
- A hand-picked basket of large stocks instead of an index fund. Every basket we built trailed the fund after dividend tax, trading costs and capital-gains tax. The gap was structural, not bad luck.
- Intraday chart-pattern setups. The average gain per trade was smaller than the taxes and fees of a single round trip.
- Order-book signals over seconds to minutes. On our first day of live order-book data, even perfect foresight of the largest moves barely covered the cost of trading them.
- Trading other public filing feeds on their own. None cleared our cost bar.
No edge, or only luck
- Reinforcement-learning agents trading on daily prices. They did not beat a placebo agent within our compute budget.
- A bigger machine-learning model to rank stocks. Its estimated gain over holding the same stocks in equal weights was below zero at every model size. The test could only have seen a large gain, so this is a closed estimate, not a proof.
- Buying after a big news-driven jump. There was no follow-through. The only positive results were the best of hundreds of variants, which is what luck looks like.
- Using commodity price shocks to predict company results. The shocks did predict reported results, but buying on them lost money even before costs.
- Writing options on the index for income. No rule we tested beat simply holding the index fund after costs and tax.
The market got there first
- Buying stocks announced for addition to a major index. They rose before the announcement, not after it.
AI that fell short of the bar
- An AI model reading company filings to forecast earnings surprises. It was right 74.2% of the time, against 72.8% for the numbers-only forecast: short of the bar.
- An AI model mapping which companies supply which, from annual reports. The pilot missed the accuracy bars we set in advance, so the full build never started.
What they have in common
Costs and taxes closed four of the twelve, and five showed no edge, or only luck. An idea has to pay for its round trip first, and in India that means securities transaction tax, stamp duty, exchange charges and GST as well as the spread.
The two AI ideas were judged like the rest: against a simpler method, or a bar fixed before the test. One idea looked good only as the best of hundreds of variants; counting every attempt is what told us it was luck.
If you can see a way to reopen one of these, or a better test for it, we would like to hear it.