Eleven hypotheses, two rounds, zero trades
In six weeks I built an autonomous research rig for Indian equities: a tick capture plane that records the full order book of fifty names every trading day, a replay engine that streams those ticks back through the live decision code, and a scoring harness that measures whether a candidate signal predicts anything at all.
Then I used it to test eleven hypotheses about intraday equities, registered across two rounds.
None of them cleared its bar. Nine failed outright and two voided themselves on their own registered conditions. No strategy was built on any of them. No capital was deployed. The rig has never placed an order with real money and still hasn’t.
That is the result I want to lead with, because the alternative — leading with the one thing that eventually passed — is how almost everyone in this field lies, mostly to themselves.
The gate came first
The rule was fixed before any of it ran: no signal reaches engine implementation until it clears the cost hurdle by a stated multiple, across at least ten captured days, spanning more than one regime family, at the horizon the strategy would actually hold.
That sentence was committed to the repository, with a date, before the data existed to test it against. Everything downstream is just execution.
The reason to fix the bar in advance is not integrity theatre. It is that a threshold chosen after seeing the results is not a threshold, it is a description. Every quantitative researcher knows this. Very few of us behave as though we do, because the moment a number comes back at 1.7× a bar you set at 2×, the mind produces an extraordinary volume of reasons why 1.7 was always the sensible bar.
What eleven failures actually looked like
They were not all the same shape, and the differences are the interesting part.
The one that looked best was luck. A flow-divergence signal came out of a five-day preview as the standout — positive at every horizon, positive on four of five days, strongest exactly where conviction was highest. It sat at 1.5–1.9× the hurdle, short of the pre-committed bar. Five more clean days were banked without ever being scored; a rehearsal was explicitly declined in order to keep them unseen. When the full ten days ran, the five new ones came in at roughly −9 bps at thirty minutes and inverted the whole result. The conviction signature flipped sign: what had been the strongest bucket became the worst.
Small frequent wins, catastrophic losses precisely where the signal was most confident. A metric with a 55–61% preview win rate was noise wearing a costume, and five unseen days caught it before a line of engine code existed.
The one that was real was unaffordable. A resting-book pressure effect measured positive on ten days out of ten, at every horizon, across sixty-four thousand events, with cleanly rising conviction buckets. It is, as far as I can tell, a true microstructure fact. It is also roughly a fifth of the size of my round-trip costs. The order book does whisper. At one basis point, to somebody with a retail fee schedule, inaudibly.
The closest miss got to 0.97× of the bar — and the bar it got to 0.97× of was the single cost hurdle, not the double-cost gate that would have promoted it. Close only counts if you decided in advance that close counts.
Why this isn’t a capital problem
The obvious objection is that all of this is a small-account problem: the fees are eating a real edge, so raise the capital and the edge appears.
I tested that, because it was my hope too. The same strategy replayed at the base capital, at four times it, and at ten times it lost about the same percentage at every level. Fixed costs shrank exactly as predicted and it changed nothing, because the binding constraint was never the fees. Below the fees sits an irreducible floor of regulatory charges and bid-ask spread that no account size removes, and the best signal measured never cleared it.
Intraday equity research here is closed. Not paused pending more money — closed, on the evidence, and written down with a date so I cannot quietly reopen it when I get bored.
What the six weeks actually bought
Not a strategy. An instrument, and a record.
The capture plane still runs every morning, because the archive is the durable asset — a decade of survivorship-free daily history with every corporate action hand-verified, plus a growing tick archive with full depth. The scoring harnesses, the replay engine, the corporate-action machinery, and the pre-registration protocol all exist and all work.
And there is now a public ledger of every hypothesis this programme has registered — across both the intraday work described here and a parallel swing track — with its pre-committed bar and its verdict. Twenty-three entries. One qualified pass, currently being tested forward on data that did not exist when its thresholds were published.
Twenty-two of those entries are dead. That is not the disclaimer. That is the portfolio.