Algorithmic Trading System

An early trend model, EMA bias with a DOL filter, looked spectacular: a clean, rising equity curve and a healthy profit factor across fifteen years. It was too good. An audit of the swing-detection logic found the culprit: it was confirming swing highs and lows using bars that hadn’t happened yet. The “edge” was the strategy quietly reading the future. Rebuilt to use only information available at the moment of each decision, the same system’s profit factor collapsed from comfortably profitable to 0.63, a loser. That failure became the most valuable result in the project: every strategy since has to pass a causal parity test before it goes anywhere near a live account.

With the validation harness in place, we ran a systematic search across strategy families and distinct market eras. The honest conclusion: no strategy held a durable profit factor above about 1.25 in every era. The strongest, most stubborn lead was a quality-filtered Donchian daily breakout, hovering around 1.15 to 1.27 depending on the period; classic setups like RSI-2 mean reversion “died honest”, with no survivorship illusion left to hide behind. Documenting negative results like these is the process, not a failure of it: it’s the line between real research and a curve-fit fantasy.

The live component runs the strongest structural idea end to end on EURUSD: wait for the Asian session’s range to sweep liquidity into the London open, then enter on a one-minute change-in-state-of-delivery, place the stop just beyond the swept swing, and target roughly 2.2R. It executes fully automatically through the MetaTrader 5 Python API, handling live ticks, order placement and position management under a fixed per-trade risk cap. It ran as a genuine live experiment, with real spread, real slippage and real fills, to pressure-test the model where a back-test simply can’t.