All Posts

Population Stability Index (PSI) measures how much a trading model's return distribution has shifted from its validated baseline, using a bin-by-bin comparison that detects gradual drift across the full range of outcomes. Borrowed from credit risk modelling, it provides interpretable thresholds and complements the KS statistic, Wasserstein Distance, and Jensen-Shannon Divergence in darwintIQ's distribution similarity assessment.

Mutual information is an information-theoretic metric that measures how much the return distribution of one period informs another. Unlike correlation, it captures non-linear statistical dependencies — making it a more sensitive tool for detecting whether a trading model's behaviour has remained structurally consistent. darwintIQ uses mutual information as part of a suite of distribution similarity metrics evaluated on a rolling basis.

The KS statistic applies the Kolmogorov-Smirnov test to compare return distributions across time periods, flagging when a model's behaviour has shifted from its validated baseline. It is a non-parametric measure that works alongside Jensen-Shannon Divergence, Wasserstein Distance, and PSI in darwintIQ's distribution similarity assessment, helping traders identify structural drift before it shows up in headline figures.

Standard deviation measures how far a trading model's returns spread around their average, revealing the consistency behind headline performance. It underpins risk-adjusted metrics including the Sharpe and Sortino Ratios and helps identify whether a model's results are genuinely reliable or deceptively volatile. In darwintIQ, it forms part of a broader performance dashboard alongside drawdown, Expected Value, and distribution metrics.

An adaptive trading system continuously updates its assessment of which models are suited to current market conditions, rather than deploying a fixed strategy based on historical optimisation. In darwintIQ, this adaptation is achieved through continuous model evaluation on a rolling window — meaning model rankings reflect what is working right now, not what worked at some point in the past.