The question
Many backtests look profitable because they ignore what trading actually costs: fees, execution, position limits, and the gap between the data a strategy was tuned on and the data it wasn't. We wanted a platform that tells us when a strategy doesn't work.
What we built
- Exact accounting. Cash, positions, prices and fees use exact-decimal arithmetic, so rounding errors can't quietly change a strategy's profit.
- Replayable history. Every market event and decision goes into an append-only journal, so any result can be replayed and audited.
- Paper trading only. Live order submission is blocked by design.
- Screening at scale. We tested strategies across 736 instruments and about 9.5 GB of historical Kraken data.
The platform did its job by rejecting strategies. Several failed our criteria once costs were included. One momentum strategy, for example, turned its portfolio over about 156 times a year, and trading costs erased its returns.
My part
I helped shape the initial research and pushed for OCaml. Its type system rejects invalid trading states before the code ever runs, so whole classes of costly bugs can't happen.
I built three of the platform's pre-trade risk checks:
| Check | Risk it prevents | How it works |
|---|---|---|
| Maximum-loss kill switch | A losing strategy keeps trading | Tracks account equity against a baseline; past the limit, it blocks new orders and cancels open ones |
| Order-size limits | One oversized order creates outsized exposure | Caps each order's quantity and dollar value, with tighter limits per strategy |
| Capital availability | Two strategies spend the same cash | Before approving an order, checks free funds net of open orders and estimated fees |
I also built the entire research frontend: a React and TypeScript terminal with Dashboard, Strategies, and Activity & Risk workspaces, dockable panels, and portfolio and performance charts.
Status
The core engine and the frontend are built. Connecting them is the next step. No real money has been traded.
What I learned
The most valuable thing a backtester can do is say no. A strategy that looks good before costs tells you very little. And a risk check is only as good as its failure mode: when in doubt, block the order.
Code available on request.