factor-backtest

Name: factor-backtest
Author: minihellboy

95stars28forks

How to install factor-backtest?

minihellboy/factorminer/factor-backtest

$npx -y skills add minihellboy/factorminer --skill factor-backtest

Installs into the current project.

Summary

Combine a factor library into a composite signal and quintile-backtest it under transaction costs — long-short return, monotonicity, turnover, and tearsheets. Use for the portfolio-level view that single-factor IC does not give. Triggers on "backtest", "composite signal", "combin

Files · 1

View on GitHub

SKILL.md

1	# Factor Backtest
2
3	A library of individually-decent factors is not a strategy. This skill combines them into one composite signal and backtests the portfolio that signal implies — the level at which transaction costs and capacity actually bite.
4
5	## Workflow
6
7	### 1. Combine and backtest
8
9	```bash
10	factorminer combine output/run1/factor_library.json \
11	--data path/to/market_data.csv \
12	--method all --fit-period train --eval-period test
13	```
14
15	- `--method` — `equal-weight`, `ic-weighted`, `orthogonal`, or `all` to compare every method.
16	- `--fit-period` — split used to fit weights / run selection (use `train`).
17	- `--eval-period` — split used to score the composite (use `test`).
18	- `--selection` — optional pre-filter: `lasso`, `stepwise`, `xgboost`, or `none`.
19	- `--top-k` — keep only the top-K factors before combining.
20
21	The report gives composite `IC Mean`, `ICIR`, `Long-Short` return, `Monotonicity`, and `Avg Turnover`.
22
23	### 2. Generate tearsheets
24
25	For the visual portfolio view — quintile returns, IC time series, correlation heatmap:
26
27	```bash
28	factorminer -o output/run1 visualize output/run1/factor_library.json \
29	--data market_data.csv --period test --tearsheet --quintile --correlation
30	```
31
32	## What to look for
33
34	- Monotonicity — quintile returns should step up Q1→Q5. A non-monotone composite is fragile regardless of headline IC.
35	- Long-short return net of turnover — high `Avg Turnover` means the gross return is optimistic; FactorMiner's transaction-cost model is what makes the net number honest.
36	- Method spread — if `orthogonal` and `equal-weight` disagree sharply, the library has redundant or unstable factors; revisit `factor-evaluation`.
37
38	## Guardrails
39
40	- Fit weights on `train`, score on `test` — never fit and score on the same split.
41	- The backtest estimates historical behavior; it is not a forward return promise. Present it as a research artifact for review.
42	- Report net-of-cost numbers as the headline; gross numbers only as context.

People who liked this also viewed

6 picks

Type

factor-backtest

95stars28forks

View on GitHub

How to install factor-backtest?

minihellboy/factorminer/factor-backtest

$npx -y skills add minihellboy/factorminer --skill factor-backtest

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Factor Backtest
2
3	A library of individually-decent factors is not a strategy. This skill combines them into one composite signal and backtests the portfolio that signal implies — the level at which transaction costs and capacity actually bite.
4
5	## Workflow
6
7	### 1. Combine and backtest
8
9	```bash
10	factorminer combine output/run1/factor_library.json \
11	--data path/to/market_data.csv \
12	--method all --fit-period train --eval-period test
13	```
14
15	- `--method` — `equal-weight`, `ic-weighted`, `orthogonal`, or `all` to compare every method.
16	- `--fit-period` — split used to fit weights / run selection (use `train`).
17	- `--eval-period` — split used to score the composite (use `test`).
18	- `--selection` — optional pre-filter: `lasso`, `stepwise`, `xgboost`, or `none`.
19	- `--top-k` — keep only the top-K factors before combining.
20
21	The report gives composite `IC Mean`, `ICIR`, `Long-Short` return, `Monotonicity`, and `Avg Turnover`.
22
23	### 2. Generate tearsheets
24
25	For the visual portfolio view — quintile returns, IC time series, correlation heatmap:
26
27	```bash
28	factorminer -o output/run1 visualize output/run1/factor_library.json \
29	--data market_data.csv --period test --tearsheet --quintile --correlation
30	```
31
32	## What to look for
33
34	- Monotonicity — quintile returns should step up Q1→Q5. A non-monotone composite is fragile regardless of headline IC.
35	- Long-short return net of turnover — high `Avg Turnover` means the gross return is optimistic; FactorMiner's transaction-cost model is what makes the net number honest.
36	- Method spread — if `orthogonal` and `equal-weight` disagree sharply, the library has redundant or unstable factors; revisit `factor-evaluation`.
37
38	## Guardrails
39
40	- Fit weights on `train`, score on `test` — never fit and score on the same split.
41	- The backtest estimates historical behavior; it is not a forward return promise. Present it as a research artifact for review.
42	- Report net-of-cost numbers as the headline; gross numbers only as context.

People who liked this also viewed

6 picks

Type