Backtesting hub
Built to separate an edge from chance.
A place to test trading ideas against recorded market data and computed dealer positioning, designed around the ways backtests usually go wrong.
In development. The research tooling described here runs internally today. The hosted hub is being built and is not open yet.
Why another backtester
A backtest is easy to run and hard to trust.
Three things break most results before anyone looks at them: information from the future leaking into the inputs, many variations tried and only the best one reported, and costs left out. None of them shows up as an error. The number just looks better than it should.
The hub treats each of these as a property of the system, not as something the researcher has to remember.
- Look ahead. A feature that quietly contains tomorrow's price.
- Selection. Twenty variations tested, one reported.
- Friction. An edge of half a tick, before a spread of one.
- Drift. History computed by different code than the live figure.
How the hub answers them
One engine for past and present
Every historical session is produced by replaying the exchange record through the same engine function that computes a figure during the session. There is no separate "research version" of a level that could differ from what was computable at the time.
Forward returns are kept apart
Outcomes live in their own table, every column is named as forward looking, and they are joined to features by timestamp, never by row position. A gap in a session cannot shift a label onto the wrong row.
Every trial is counted
Each run is written to a registry by the tool, not typed in by the person who wants it to be good. Tests are grouped into families that bear on one question, and significance is adjusted for how often that family has been tried. An idea that was never tested shows as never tested.
Costs come first
You set the spread and commission assumptions once. Every result is shown net of them next to the raw figure, so an effect smaller than the cost of trading it is visible as such.
Days are checked before they count
A session enters the research store only if it is complete. Sessions with gaps are marked, repaired from the historical record where possible, and otherwise excluded with the reason stated.
Independent recomputation
Any stored session can be recomputed from the price series alone by a separate program that shares no code with the pipeline, and compared against the store.
What you work with
Inputs
Market data and positioning
Bars, trades, quotes, open interest and the computed exposure tables for ES and NQ.
Coverage.
Method
Events and hypotheses
Define an event such as a level touch, a regime change or a flow threshold, state the hypothesis before running it, and measure what followed at fixed horizons.
Output
Results you can check yourself
Sample size, effect, cost adjusted effect, adjusted significance, and the code that produced them. Exportable.
Or hand it to Mason ยท in development
Rather describe the test than build it?
Ask him: Test whether a break of the call wall before 14:00 holds into the close, net of costs.
He does the digging and hands back the result with the code and the sample size.
Meet Mason โ
What it is not
The hub does not place orders, send signals or connect to a broker. It is a research tool. What you do with a result is your decision, and a result that holds up in a backtest can still fail in the market.