Skip to content

Signal persistence

Signal persistence measures how scores change before any portfolio is constructed. It is independent of forward-return horizons, delay, targets, volatility warmup and return availability. These diagnostics do not compute positions, traded notional, costs or P&L.

Available diagnostics

Every new evaluation computes result.persistence. For a runnable example and settings, see measure signal persistence. The definitions below explain how missing observations and changing asset sets affect the results.

An eligible score has a finite score and a price at its own timestamp. Each comparison uses endpoints exactly lag rows apart on the price grid. It never drops missing rows before shifting. A lag of 5 compares endpoints five bars apart, even if an interior bar is missing; it neither measures the intervening path nor sums five one-bar changes. A score that moves from positive to negative and back between the endpoints can have zero endpoint change. Increasing the lag therefore changes the comparison, rather than accumulating turnover. Both endpoints must be inside the evaluation window; no score history before start is read. All timestamps through the last bar can contribute, even when they have no future-return label.

Cross-sectional rank persistence

Rank all eligible scores independently at each timestamp, using average ranks for ties. Compute Pearson correlation of those ranks on the assets common to the two endpoints. This follows Alphalens Reloaded's rank-autocorrelation convention; with changing universes it need not equal re-ranking only the common assets. The common sample needs min_assets and varying ranks at both endpoints. Values range from −1 (reversed ordering) to 1 (preserved ordering). High persistence does not establish low dollar turnover: weights, leverage and volatility scaling can still change.

Bucket membership

Assign scores at each timestamp to ceil(quantiles * average_rank / n_eligible). Ties stay together and some buckets can be empty. Each endpoint must have at least min_assets and distinct scores. Following Alphalens' membership comparison, for bucket Q, entry fraction is |Q_now − Q_before| / |Q_now|. A value of 0.2 means 20% of current members were absent from the prior bucket. An empty current bucket is undefined; an empty prior bucket at an otherwise eligible timestamp gives 1 for a nonempty current bucket.

Entry and exit counts are both reported: changes in bucket size can make them different. Counts of bucket entrants absent from the prior eligible universe and exits absent from the current universe distinguish availability changes from movement between buckets. These are observed score/price availability changes, not confirmed listings, delistings or forced trades. Do not interpret the first lag rows' structural counts as actual entries: their comparison status is boundary_excluded because there is no earlier endpoint in the window.

Time-series signal changes

For each asset, let S contain valid endpoint pairs (a, b) at the chosen lag:

\[ \begin{aligned} \text{normalised change} &= \frac{\sum_S |a - b|}{\sum_S (|a| + |b|)/2} \\ \text{sign-change fraction} &= \frac{\#_S\{\operatorname{sign}(a)\operatorname{sign}(b) = -1\}}{|S|} \\ \text{zero-transition fraction} &= \frac{\#_S\{\text{exactly one endpoint equals zero}\}}{|S|} \end{aligned} \]

Normalised change is in multiples of typical absolute signal over the lag, without annualisation. Its range is 0 to 2, and positive constant rescaling leaves it unchanged. It is a ratio of sums, not an average of individual pair ratios. It depends on the signal's zero point. Constant nonzero scores give zero; all-zero paired scores give zero_signal for normalised change while both transition fractions are valid zeros. Pairs with two zero scores are included in transition denominators. Too few pairs for persistence.min_pairs gives insufficient_pairs for all three statistics. The default minimum is 2, a descriptive availability rule rather than a statistical significance threshold.

Aggregation and interpretation

Cross-sectional and bucket summary means weight valid timestamps equally; time-series summaries weight assets with available statistics equally. They do not pool observations across assets. The summary includes contributing and candidate counts. No confidence intervals or significance claims are attached. Fractions are stored as decimals, lags as integer bars, and no quantity is annualised. These diagnostics measure neither return nor portfolio turnover.

The older time_series.assets.signal_autocorr_lag1 field remains unchanged for compatibility. It uses that table's horizon-maturity sample and measures an asset's own score autocorrelation; the cross-sectional rank statistic measures ordering across assets and uses all eligible endpoint pairs independently of return horizons.

Why score changes are not turnover

A score can stay constant while volatility scaling changes the position. Holdings also drift as prices move, so the trade needed to reach a target depends on the holdings before the trade. Neither effect is measured by score persistence. Lim, Zohren and Roberts (2019, §VI) compute changes in volatility-scaled exposure, which depend on both the score and the volatility estimate. Boyd et al. (2017, §§2.2–2.4) distinguish trades, transaction costs and holding costs in a portfolio accounting model. These are quantities that score changes alone cannot supply; costs cannot be subtracted from IC.

The normalised-change statistic above is alphatonic's descriptive convention. Its symmetric endpoint scale and lack of annualisation should be retained when comparing results. The cited trading papers motivate examining exposure changes; they do not validate this statistic as a cost estimator.

References