ic_summary¶

Defined in fynance.metrics

ic_summary(pred, real, method='spearman')[source]

Summary statistics of the per-bar cross-sectional Information Coefficient.

Reduces a (T, N) panel to the headline numbers of a factor tear-sheet. The per-bar cross-sectional IC (one value per bar, across the N assets) is computed with fynance.metrics.information_coefficient, then summarized. icir (the IC information ratio) and t_stat gauge whether the mean IC is distinguishable from zero given its bar-to-bar variability.

Alignment. pred[t] is the score known at t and real[t] the outcome realized after t.

Parameters:

prednp.ndarray[dtype, ndim=2]: Factor/score panel (T, N).
realnp.ndarray[dtype, ndim=2]: Aligned forward-return panel (T, N).
method{‘spearman’, ‘pearson’}, optional: Correlation used for the IC (default 'spearman').

Returns:

dict of str to float: mean_ic (mean per-bar IC), icir (mean / std of the per-bar IC), t_stat (icir * sqrt(n_bars)), hit_rate (share of bars with a positive IC) and n_bars (number of bars with a finite IC). icir and t_stat are nan when the IC has zero variance.

See also

fynance.metrics.information_coefficient, roll_information_coefficient

Notes

With \(IC_t\) the per-bar IC over the \(n\) bars with a finite value,

\[\mathrm{ICIR} = \frac{\overline{IC}}{\sigma_{IC}}, \qquad t = \mathrm{ICIR}\,\sqrt{n}\]

where \(\sigma_{IC}\) is the sample standard deviation (ddof=1).

Examples

A factor equal to the realized outcome has a perfect per-bar IC:

>>> import numpy as np
>>> rng = np.random.default_rng(1)
>>> real = rng.normal(size=(100, 10))
>>> pred = real.copy()
>>> s = ic_summary(pred, real)
>>> round(s['mean_ic'], 6)
1.0
>>> round(s['hit_rate'], 2)
1.0
>>> s['n_bars']
100