openquant.sample_weights

Compiled from Rust (crates/pyopenquant). Signatures and types are from the shipped .pyi stubs; descriptions are the functions’ own docstrings, as help() shows them. All Python modules.

Functions

get_weights_by_return

get_weights_by_return(
    events: Sequence[tuple[str, str, float]],
    close_timestamps: Sequence[str],
    close_prices: Sequence[float],
) -> list[tuple[str, float]]

Sample weights by return attribution (AFML section 4.6, Snippet 4.10).

For each event i, w_i = |sum_{t in [start_i, end_i]} r_t / c_t|, where r_t is the log return arriving at bar t (computed over the whole series, so the return arriving at start_i is included) and c_t is the number of events whose span contains bar t. The weights are then scaled to sum to the number of events (mean 1), unless they are all zero. Spans are inclusive and matched to the close series by exact timestamp. Concurrency is computed over the events passed in, so compute weights on the training fold only. A single large move (a gap or bad print) can take most of the weight.

Parameters

events : list[tuple[str, str, float]]

(start, end, label) per event, timestamps as "%Y-%m-%d %H:%M:%S" (an optional fractional second is accepted). The label is ignored.

close_timestamps : list[str]

Bar timestamps in ascending order, in the same format.

close_prices : list[float]

Strictly positive close prices (a non-positive price gives non-finite weights).

Returns

list[tuple[str, float]]

One (event_start, weight) pair per event, in input order.

Raises

ValueError

If a timestamp does not parse, the close timestamps and prices differ in length, or an event's end precedes its start.

get_weights_by_time_decay

get_weights_by_time_decay(
    events: Sequence[tuple[str, str, float]],
    close_timestamps: Sequence[str],
    close_prices: Sequence[float],
    decay: float,
) -> list[tuple[str, float]]

Sample weights by time decay along cumulative uniqueness (AFML section 4.7, Snippet 4.11).

Each event's average uniqueness is the mean of 1 / c_t over the close bars in its inclusive span (c_t = number of events containing bar t); an event covering no bar has uniqueness 0. With x_i the cumulative uniqueness in start order and X its total, the weight is max(0, a + b x_i) with b = (1 - decay) / X for decay >= 0, b = 1 / ((decay + 1) X) for decay < 0, and a = 1 - b X, so the newest event has weight 1. decay = 1 is no decay, 0 < decay < 1 decays linearly toward decay, 0 toward 0, and -1 < decay < 0 zeroes the oldest -decay fraction. The weights are not rescaled and exclude return attribution (AFML suggests multiplying the two). Close prices are not used, only their timestamps.

Parameters

events : list[tuple[str, str, float]]

(start, end, label) per event, timestamps as "%Y-%m-%d %H:%M:%S" (an optional fractional second is accepted). The label is ignored.

close_timestamps : list[str]

Bar timestamps in ascending order, in the same format.

close_prices : list[float]

Close prices; only their count is checked.

decay : float

Decay parameter; must be in (-1, 1].

Returns

list[tuple[str, float]]

One (event_start, weight) pair per event, in input order.

Raises

ValueError

If a timestamp does not parse, the close timestamps and prices differ in length, or an event's end precedes its start.