Repository navigation
amihud_illiquidity nulls a whole window for one untraded bar #31
Description
features/microstructure/amihud_illiquidity.py:198-199 nulls the bar where dollar volume is zero
and then takes rolling_mean(period) over that series:
.otherwise(None)
.rolling_mean(period)Polars' rolling_mean propagates nulls, so one untraded bar nulls the whole following window,
not just itself.
Measured
Called with period=30 on a NASDAQ-100 minute panel, 2020-01-02 to 2020-03-31, 2,074,372 rows:
| Column | Null share |
|---|---|
illiq |
9.08% |
dollar_vol (the underlying, null only on the bar itself) |
0.64% |
0.64% of bars are untraded and 9.08% of rows carry no Amihud value: a ~14x amplification, which is
period acting on isolated gaps.
Why this is a defect and not a convention choice
Amihud (2002) defines ILLIQ as the average of |R_d| / DVOL_d over the D days the stock traded in
the period - the sum is over traded days and the denominator is their count. The implementation
instead requires every bar in the window to have traded.
On daily equity data the two agree, because a listed name trades every day. On intraday data they do
not, which is where the 14x appears.
Suggested direction
Give the rolling mean min_periods (or skip nulls) so the average is taken over the bars in the
window that traded, matching the definition. That is a behaviour change for any caller with gaps, so
the docstring should state which convention is used.
Activity