Eight strategies, one number
A thread on X published eight backtests that all landed within half a point of each other, then called one of them the winner. What the table actually says.
This week a long thread went round X about building a trading system with AI. Somewhere in the middle of it, almost as an aside, the author posted a table. Eight textbook strategies, each run out of sample with costs on. Eight results, all between minus 17.8 percent and minus 18.2 percent.
Then the next line said: build on the winning strategy.
Read the spread first
Eight different rule sets. Different indicators, different entries, different exits. And they land within 0.4 of a point of one another.
When eight different things produce one number, the things are not producing the number. Something else is. In this case, the window of time they were run over and the costs they were charged. The rules were decoration on a loss that was already decided before any rule was written.
The author ran the honest test, got the honest answer, and then treated the noise between eight losers as a ranking.
That is the whole lesson, and it is a lesson about reading, not about trading.
We tested one of them anyway
On this desk a claim does not get an opinion. It gets a test with the verdict written before the run. So we took the strategy from the thread's own screenshot, a classic oversold entry with a trend filter, wrote down what result would confirm the table and what would not, and ran it twice, on two different data sources, with the same code.
It finished below simply holding the coin in every window we tried. Not because the costs were cruel. With costs switched off it still lost. The rule itself was wrong-footed more often than it was right, and the fees finished the job.
The version of it that trades often lost more. The version that trades rarely barely traded at all. Neither version beat doing nothing.
What the table is good for
It is the best teaching exhibit I have seen this year, and I am putting it in front of every student I have.
Not because the strategies are bad. Because the reading was. A table like that is a gift. It tells you, for free, that the thing you are about to spend a month tuning is not where the number comes from. The next step is not to pick the top row. It is to ask what all eight rows have in common.
Three habits that would have caught it
Write the verdict before the run. If the rule is below holding after costs, it is rejected, and there is no second run with a nicer number.
Put holding on the same chart. A strategy that makes money in a market that went up has not told you anything yet.
Charge the real round trip, in and out, at the venue you actually use. The thread's table used a cost that fits one exchange and flatters every other. On the venues most of our students trade, the same strategy loses more than twice as much.
None of that is clever. All of it is the difference between a number and a story about a number.
What we did with it
The exact parameter names from the author's script went into our own tools, because those were correct and useful. The table went into the course as an exercise. The architecture around it, which fires trades on a probability nobody measured, did not go anywhere.
Take the gift. Leave the ranking.