quantskills/agent-quantspace/compute

Description

Compute derived values from OHLCV data. This skill is organized in reusable, strategy-agnostic layers:

Also: label makers (label_maker.py), forward adjustment from stored adjustment factors (adjust.py), and the Factor wrapper.

Shared market-structure helpers used by strategy domains:

• skills.compute.resample.resample_to_5m(df_1m) — A-share 1m eob OHLCV to 5m eob bars without crossing lunch break; keeps 09:31-11:30 and 13:01-15:00, removes zero-volume rows, and preserves OHLCV aggregation.
• skills.compute.regime.split_by_regime(df, regimes=None) — lithium-cycle date slicing for DatetimeIndex or MultiIndex inputs.

Strategy-specific factors and feature engineering live in strategies/, not here — except generic single-instrument OHLCV transforms such as log-difference grids, which belong in skills.compute.features. The factor-mining workflow may generate new Python factor functions directly in strategies/<domain>/mined_factors/; they only need to satisfy the Factor callable contract below. They do not need to be added to skills.compute.indicators or to any allowlist.

Prerequisites

• Python: pandas, numpy.
• Imports:
• from skills.compute.indicators import trend_score, rsi, er
• from skills.compute.utils import safe_divide, calculate_atr
• from skills.compute.wrappers import Factor
• from skills.compute.resample import resample_to_5m
• from skills.compute.regime import split_by_regime
• from skills.compute.features import make_logdiff_features, default_logdiff_shifts

API Reference

Math Utilities (skills.compute.utils)

from skills.compute.utils import safe_divide, rolling_zscore, calculate_atr, clip_outliers, round_away_from_zero

6 public functions: safe_divide, rolling_zscore, rolling_regression_vectorized, calculate_atr, clip_outliers, round_away_from_zero, plus private helpers _weighted_polyfit_coefficients, _rolling_linear_regression, _scalar_kalman_smoother.

Universal Indicators (skills.compute.indicators)

from skills.compute.indicators import trend_score, rsi, er, supertrend

19 public functions organized by category:

• Price/Momentum: roc, ma, daily_return, ma_cross, price_above_ma, bias_momentum, mom_skip
• Trend: trend_score, trend_score_v2, trend_score_v2_skip, supertrend, donchian_channel
• Volume: orb_relvol
• Efficiency: er
• Oscillators: cci, slowkdj, williams_r, rsi, rsi_divergence

Factor wrapper (skills.compute.wrappers)

from skills.compute.wrappers import Factor
from skills.compute.indicators import trend_score_v2

• __init__(func: Callable, **params) — binds callable and defaults; name from func.__name__ and params.
• calculate(data: pd.DataFrame, *, dropna: bool = True) -> pd.Series — per-symbol apply with index contract checks. Pass dropna=False to keep warm-up NaNs and the full index (factor_mining always does).
• cal_df(data, *, dropna: bool = True) -> pd.DataFrame — wide pivot with eob index and one column per symbol; respects the same dropna flag. Not an authoritative research result.

Factor never persists files. Artifact writes belong to factor_mining store adapters.

Contract for `func`: first argument is a single-symbol DataFrame with a one-level eob DatetimeIndex and required OHLCV columns; return a real numeric Series whose index equals the input index item-for-item.

Exit / risk filters (cross-sectional package)

from skills.strategy.cross_sectional.exits import (
    gap_down_filter,
    vol_spike_filter,
    drawdown_from_high_filter,
)

Panel-level filters return a MultiIndex Series; use with ExitFilterConfig and condition in the cross-sectional backtester (see strategies/cross_sectional/STRATEGY.md).

Recipes

1. Panel factor

import pandas as pd
from skills.compute.wrappers import Factor
from skills.compute.indicators import roc, trend_score_v2
f = Factor(trend_score_v2, period=24)
scores = f.calculate(data)  # data: MultiIndex (symbol, eob)
f2 = Factor(roc, period=60)
roc_df = f2.cal_df(data)

2. Direct single-symbol call Slice one symbol, index by eob → roc(sym_df, period=20).

3. Apply a single-symbol indicator to a panel

from skills.compute.indicators import trend_score
from skills.compute.wrappers import Factor
factor = Factor(trend_score, period=25)
scores = factor.calculate(panel)

4. Exit filter in a backtest Pass {'func': drawdown_from_high_filter, 'kwargs': {...}, 'condition': lambda x: x < 0.1} in exit_filters on ModularBacktester (see strategy domain doc).

OHLCV Features (skills.compute.features)

from skills.compute.features import (
    default_logdiff_shifts,
    make_logdiff_features,
    make_logdiff_panel_features,
)

• make_logdiff_features(bars, *, factors=..., lags=..., shifts=None, lookback=5) — single-symbol OHLC log-difference grid. Default: 4×4 factor pairs × 9 lags × 10 shifts = 1440 columns when lookback=5. Non-positive prices become NaN instead of -inf.
• make_logdiff_panel_features(panel, ...) — same grid applied per symbol on a (symbol, eob) panel, non-finite rows dropped, so the result is model-ready. Build once and reuse across horizons/models.
• default_logdiff_shifts(lookback) — tuple 0..lookback-1 followed by 5, 10, ..., lookback*5 (not deduplicated).
• No duplicate log-difference implementation under strategies/.

Label Generation (skills.compute.label_maker)

from skills.compute.label_maker import ForwardReturnLabelMaker, TripleBarrierLabelMaker

Public supervised-learning labels:

• ForwardReturnLabelMaker: forward-return threshold labels.
• TripleBarrierLabelMaker: AFML-style volatility-scaled triple-barrier labels.

CLI command reference

<!-- BEGIN GENERATED: qs commands -->

Transport: --input-json FILE|-, --preview N, and --log-level LEVEL; flags override JSON. Use --print-template [FIELD] for fillable templates.

Shared parameter groups

PanelSource

OutputOptions

uv run qs compute indicators list

List public indicators with signatures.

Priority: P0. Side effects: read-only.

Output data fields: indicators.

Inspect schemas: uv run qs schema compute.indicators.list (add --part input|output or --file-field FIELD).

uv run qs compute indicators show

Show one indicator's signature, defaults, and docstring.

Priority: P0. Side effects: read-only.

Positional: name.

Output data fields: name, signature, defaults, docstring.

Inspect schemas: uv run qs schema compute.indicators.show (add --part input|output or --file-field FIELD).

uv run qs compute factor

Calculate an indicator or module:name factor on a panel.

Priority: P0. Side effects: writes-output.

Shared parameters: PanelSource, OutputOptions.

Output data fields: name, rows, symbols, start, end, nan_ratio, output.

Inspect schemas: uv run qs schema compute.factor (add --part input|output or --file-field FIELD).

uv run qs compute features

Build model-ready OHLC log-difference features for a panel.

Priority: P1. Side effects: writes-output.

Shared parameters: PanelSource, OutputOptions.

Output data fields: rows, features, symbols, start, end, output.

Inspect schemas: uv run qs schema compute.features (add --part input|output or --file-field FIELD).

uv run qs compute labels forward-return

Create forward-return threshold labels for one symbol.

Priority: P1. Side effects: writes-output.

Shared parameters: OutputOptions.

Output data fields: rows, columns, output.

Inspect schemas: uv run qs schema compute.labels.forward-return (add --part input|output or --file-field FIELD).

uv run qs compute labels triple-barrier

Create triple-barrier labels for one symbol.

Priority: P1. Side effects: writes-output.

Shared parameters: OutputOptions.

Output data fields: rows, columns, output.