A research platform for algorithmic trading

The classic failure of quantitative research is a backtest that shows profit which live trading never reproduces. The causes are almost always infrastructural: the test includes instruments that cannot actually be traded, the fee is understated by half, and the symbol universe at the moment of the trade differs from today's.

On one of our own harnesses this was expensive: part of the results turned out to be fiction, because the sample included instruments unavailable for real execution.

We rebuilt the pipeline around one principle — the test may never know more than the trading system would have known at that moment.

We added a tradability check for every symbol against the exchange's allow-list; deny-lists do not work here, because new instruments appear faster than anyone adds them. Of 747 perpetual contracts, 419 proved actually tradable, and those form the test universe.

We replaced the optimistic fee with the real taker rate and added a parity test: the same strategy runs against the exact instrument universe, and a result is accepted only if the confidence interval on returns sits above zero.