Overview

Axis-agnostic diagnostics that operate on any (date, value) series produced by an upstream cell metric — information coefficient (IC) time series, \(\beta\) time series, CAAR time series, an external factor return, etc. They do not care which cell produced the series.

Not the same as DataStructure.TIMESERIES

DataStructure.TIMESERIES is the dispatch structure for n_assets == 1 (set on EvaluationResult.cell); the metrics on this page run on a (date, value) series regardless of which structure produced it.

For the single-asset dense evaluate() path, use predictive_beta. It consumes the long panel directly and tests the predictive-regression slope with Newey-West HAC inference.

Metric	Page
Out-of-sample decay across multiple splits	oos
Theil-Sen monotonic trend with augmented Dickey-Fuller (ADF) persistence flag	trend
Sign-consistency / hit rate of the series	positive_rate

All three are descriptive robustness diagnostics. Use them against the IC series from ic.compute_ic, the CAAR series from caar.compute_caar, or any externally constructed (date, value) DataFrame to layer decay, drift, and sign-stability checks on top of the cell's inferential test.

directional_hit_rate sits next to these sign diagnostics in the navigation, but it is not a (date, value) helper: it consumes the long panel (date, asset_id, factor, forward_return) and tests whether sign(factor) predicts sign(forward_return).

directional_pair_accuracy is also panel-input, but it compares same-date asset pairs for rank ordering and remains descriptive (p_value=None).