Every dataset is stored as it was first observed, not as it was later corrected. The platform retains revision history so a signal can be evaluated against the information actually available on any given date.
Survivorship-bias-free universes are constructed from point-in-time membership. Delisted, renamed, and suspended securities remain in the panel; corporate actions, identifier mappings, and vendor normalisation are versioned and replayed on their effective dates. There is no lookahead, and no future information can leak into the as-of view.
Cost-aware & event-driven
- Event-driven matching against point-in-time order books and reference prices.
- Transaction costs modeled explicitly — impact, spread, and slippage — rather than deducted after the fact.
- Turnover is a measured output of the simulation, not an input assumption.
No lookahead, no hindsight
- Signals are evaluated only against information available at the decision timestamp.
- Corporate actions and restatements are applied on their effective as-of date.
- Universe membership is re-evaluated each rebalance, not inherited from the current snapshot.
Every run is a commit
- Data version, code revision, and parameters are captured with each result.
- Any backtest can be replayed to the exact same state, for audit or for review.
- Results are diffable, so a changed assumption is visible, not silent.
Against overfitting
- Walk-forward and holdout windows are first-class objects, not conventions.
- Multiplicity in factor selection is tracked, so in-sample Sharpe cannot pose as a forecast.
- Degradation between in-sample and out-of-sample is measured and surfaced, not hidden.
Risk is reported the way a systematic manager needs it: factor exposure, drawdown analytics, and regime conditioning — all recomputed from the same data as the backtest.
Drawdown is decomposed into contributions by factor, sector, and regime rather than reported as a single number. Exposures are computed from the same point-in-time holdings as the backtest, and regime shifts feed the risk budget directly rather than being noticed after a loss.
Execution is where the simulation meets the book. The platform models schedule, impact, and the price of urgency so an implementation shortfall can be studied before it is paid.
Timing studies compare urgency against patience in the same book, and capacity estimates are derived from measured market impact rather than assumed from a turnover rule.
| Metric | Definition | Value |
|---|---|---|
| Sharpe ratio | Excess return per unit of volatility, annualized. | — |
| Sortino ratio | Excess return per unit of downside deviation. | — |
| Max drawdown | Largest peak-to-trough decline over the window. | — |
| Calmar ratio | Annualized return divided by max drawdown. | — |
| Win rate | Share of periods with positive contribution. | — |
| Annual turnover | Gross notional traded as a share of capital. | — |
| Capacity estimate | Capital at which measured impact erodes the edge. | — |
| Correlation to benchmark | Return correlation to the stated reference. | — |