Journey so far
How the strategy got from one person tapping buttons to a live execution engine with ~7 years of validated backtest. Honest about the parts that didn't work out. Still ongoing.
Five years of discretionary trading. Pattern recognition by eye, gut-feel exits, scattered results. The setups felt real; the execution wasn't.
Broker accounts, it turned out, have programmatic interfaces. So I translated the discretionary rules into code: setup detection, entry conditions, stop placement. For the first time the manual edge had a written-down definition, and a computer that could act on it without flinching at the screen.
Pulled historical bars. Replayed the strategy. The numbers were honest. Smaller than ego had hoped, but real.
A one-click 'find today's movers' became automatic, so the universe stopped being 'stocks I noticed on social.' Then a classifier for the market state around each setup, because different conditions call for different handling. It started to look like a real system.
Brought scanner, strategy, and broker together. Real-time, end-to-end, paper-trading every market day. Watching it work live changed how I think about edge.
Enforcing every quality filter was killing real entries. Switched filters to observation-only mode. Let the data accumulate before deciding what to enforce.
Multiple variants tested across five years of history, with a tiered framework for comparing risk-adjusted versus aggressive cuts. This is where the big backtest number first appeared. It is also where, much later, that number would unravel.
I harvested raw tick data around every critical moment and measured what idealized fills had hidden: spread, slippage, stop wick-through. Realistic costs took a large chunk of the return, painful but honest. The fix was not smarter exits but better entries: a handful of factors marked the losing trades, and skipping those entries lifted the realistic return and cut the drawdown more than any stop change. And because a handful of cherry-picked factors is exactly how you fool yourself, each one had to hold in years it was never chosen on before I trusted it.
Then a deeper audit found something worse than costs: the backtest had been entering trades just before each name officially qualified as a mover, buying the bounce with information it would not have had in real time. That single bug had inflated the headline by an order of magnitude. I killed it, threw out every number it had touched, and made a causality check mandatory on every backtest after. The edge was still there, just far smaller and finally honest.
Hand-tuning one strategy, one filter at a time, was slow and easy to fool myself with, so I changed the method. A machine-learning model now brute-forces the search across the whole feature space for what separates a winning trade from a losing one, while I steer it with intuition about where to look: a scout, not the trader. Every candidate it surfaces runs the same gauntlet, hold up in untrained years, survive having each piece stripped away, still pay after real spread and slippage, run live at size. Most do not survive. The ones that do are smaller than the fantasy but real, and they reach well beyond the original intraday momentum setup. Still ongoing.
Story continues below.
Roadmap
Hover a stop on the road to see what happens there.
Stage 1 / Operational
Live paper trading.
Daily paper trading on the live market-movers universe. Real entries scored against the held-to-peak baseline. The operational foundation; live every market day.
Stage 2 / You are here
Realistic execution.
Tick-level fill modeling: spread, slippage, and stop wick-through measured per trade against the historical tape. Entry quality, not stop tinkering, was the missing piece.
Stage 3
Reach bigger waters.
The same playbook, applied to a broader, more liquid universe.
Stage 4 / Destination
Multi-strategy automation.
Concurrent strategies running unattended on shared infrastructure. The endpoint of the route.