Market structure research
The research side of a systematic futures strategy. The tooling used to develop, validate and see a market structure model, kept strictly apart from anything live.
The constraint
A strategy is only as trustworthy as the backtest that validated it, and a backtest can lie in a hundred quiet ways. The worst one is look ahead: the simulator sees a bar the strategy could not have seen, the equity curve looks wonderful, and the first live week takes it all back.
What most people would have built
A notebook that loops over a dataframe. It runs in an afternoon, it is impossible to audit, and every parameter change is a new chance for the same bug to come back.
Walk forward, across instruments in parallel
Parameters are tuned on an in sample window and scored on the window after it, then the pair slides forward. The whole thing runs across many instruments at once, so a result that holds on one contract and collapses on the next is visible before anyone believes it.
A locked case harness
A set of cases with known answers is committed alongside the engine. Any change to the engine that moves those answers fails the build. The harness exists so that a refactor cannot quietly change what the strategy sees.
Seeing what the engine saw
The viewer is a plain web page with no build step. It renders the engine's structural reads on top of the bars, zone by zone, so a person can check the model's eye against their own. Most research tooling shows you a curve. This shows you the reasons.
Honest curves
Every equity curve is labelled with what produced it. In sample and out of sample are never on the same axis without saying so, and a curve from a run with a known gap in the data carries the gap.
What broke
An earlier batch backtester had a subtle look ahead bias that inflated results. Finding it is why the live system's simulator was rebuilt so strategies receive only a read only buffer of past bars, which makes look ahead structurally impossible rather than merely discouraged.
What it taught us
Being able to see what the engine saw is worth more than another statistic about it. A number tells you that something happened. A picture of the reads tells you whether you should believe it.