Eleven intraday ideas, and the cost floor none of them could clear
Over six weeks I built a setup for testing trading ideas on Indian stocks. It records the full order book for fifty stocks every trading day, it can replay those recordings through the same code that would have traded them, and it can score whether an idea predicts anything.
Then I used it to test eleven ideas about intraday trading, in two batches.
None of them passed. Nine failed outright and two were thrown out because the test itself turned out to be broken. I did not build a strategy on any of them. I did not use any real money. The setup has still never placed a real order.
I want to lead with that, because leading with the one thing that eventually worked is how most people in this field end up fooling themselves.
I set the pass mark first
The rule was fixed before I ran anything. An idea does not get built into a strategy until it beats my trading costs by a stated multiple, across at least ten recorded days, in more than one kind of market, at the time horizon I would actually hold it for.
I put that sentence in the repository, with a date, before I had the data to test it against. Everything after that was just doing the work.
The reason to fix the mark in advance is simple. If a number comes back at 1.7 times my costs when I said 2 times, my brain will immediately start explaining why 1.7 was always the sensible number. It is very good at this. Everyone knows this happens and almost nobody guards against it.
What the eleven failures looked like
They were not all the same, and the differences are the interesting bit.
The best-looking one was luck. An idea about order flow came out of a five-day preview looking strong. It worked at every time horizon, on four of the five days, and it worked best exactly where the signal was strongest. It reached 1.5 to 1.9 times my costs, short of the 2 times I had written down.
I had five more recorded days sitting there that I had never looked at. I could have taken a quick look. I did not. When I finally ran all ten days, those five unseen days came in around minus 9 basis points and turned the whole thing upside down. The cases where the signal was strongest went from being the best to being the worst.
Lots of small wins, and a few very large losses right where I had been most confident. An idea that won on 55 to 61 percent of days in the preview was noise in a costume. Five days I had not peeked at, and a pass mark I could not move, caught it before I wrote a single line of strategy code.
One of them was real and I still cannot use it. The pressure from resting buy and sell orders in the order book does predict the next few minutes. It worked on all ten days out of ten, across sixty-four thousand events, and it got better as the imbalance got bigger. As far as I can tell it is a true fact about how the market works.
Depending on account size it is somewhere between a twentieth and a third of what one round trip costs me. So it is real, it is free to observe, and I cannot trade it.
The closest miss got to 0.97. And that was 0.97 times a single unit of my trading cost, not the double that would have let it through. Close only counts if you decided in advance that close counts.
Here is all of it on one chart. The bars are what each signal actually measured. The solid line is what one round trip costs me, and the dashed line is the pass mark I had set at twice that.
Nothing reaches the solid line, so nothing was worth trading even once. That is the whole result, and it is why the next section matters.
More money would not fix this
The obvious objection is that this is a small-account problem. The fees are eating a real edge, so put in more money and the edge shows up.
I tested that, because I was hoping for it too. I replayed the same strategy at my base account size, at four times that, and at ten times that. It lost about the same percentage every time. The fixed costs shrank exactly as expected and it made no difference, because the fees were never the real problem. Underneath the fees there is a floor made of government charges and the gap between buy and sell prices. No account size removes that floor, and the best signal I measured never got over it.
So I have closed intraday research. Not paused until I have more money. Closed, based on what I measured, and written down with a date so I cannot quietly reopen it in a few months when I am bored.
What six weeks actually bought
Not a strategy. Equipment, and a record.
The recording keeps running every morning, because the data is the part that lasts: ten years of daily history that still contains the companies that went bust, plus a growing archive of order book recordings. The scoring code, the replay engine, the corporate action handling and the written test plans all exist and all work.
And there is now a public list of every idea I have tested, across this intraday work and a separate swing trading track, with the pass mark it had to beat and what happened. Twenty-three entries. One passed, and it is now being tested forward on data that did not exist when I set its pass marks.
Twenty-two of those entries are finished. I put them at the top of the page rather than the bottom, because they are the work.