Bounded (tanh) transform of a block's outcomes, robustly standardised

A bounded representation: tanh((y - median(y)) / s), where s is a robust scale (the MAD, falling back to the SD and then to 1 when those are zero).

Usage

riposte_tanh(y)

Arguments

y

numeric vector (a single block's outcomes).

Value

numeric vector the same length as y.

Details

The standardisation matters. Raw tanh(y) saturates to +/-1 once |y| exceeds about 3, so on an outcome measured in dollars or raw test points every value maps to +/-1 and the representation carries no information. Centring and scaling first keeps the transform sensitive in the body of each block's distribution while still bounding the influence of extreme values. To reproduce the proposal's raw tanh(y) exactly, use riposte_reps_proposal().

This transform is not in either built-in set: riposte_reps_default() uses riposte_huber() in the sixth place. Add it to a list of representations to use it.