Knowledge center · Validation & statistics
Population Stability Index (PSI) — distribution drift
How much the shape of the score distribution changed between the first and second half of the test window.
Formula
PSI = Σ (q_i − p_i) × ln(q_i / p_i), over score-histogram buckets i (p = first half, q = second half)
Interpretation
A model whose output distribution has drifted materially between two halves of the same backtest may be behaving differently now than when it was validated, even if the average IC over the full period still looks fine.
What good looks like
PSI below 0.10 — the score distribution is stable across the window.
What bad looks like
PSI above 0.25 is a standard industry threshold for meaningful drift, worth investigating before trusting the model's current behavior.
Typical range
0.00–0.10 stable, 0.10–0.25 moderate shift worth watching, above 0.25 significant drift.
How OmniSignal uses it
IC and Sharpe are averages over the whole window and can hide a model that has quietly changed behavior partway through — PSI is the direct test for that.
References
PSI is standard in credit-risk model monitoring (e.g., Basel model-validation guidance) and is used the same way here for a scoring model
Related concepts