European Equity Factor Research

A Python research pipeline for constructing equity factors, studying cross-sectional information coefficients, and comparing simple long-only portfolio scores.

Start with: factor engineering · alpha research · saved diagnostic charts.

Research workflow

Stage	Entry point	Purpose
Data collection	1_FETCH_DATA.py	Download adjusted stock and country-index prices using yfinance.
Factor engineering	2_DATA_ENG.py	Build momentum, reversal, volatility, liquidity, and beta features; export panels and coverage diagnostics.
Exploration	3_EDA.py	Explore the factor data and universe.
Alpha research	4_ALPHA_RESEARCH.py	Neutralize and standardize signals, estimate ICs, construct scores, and produce portfolio diagnostics.

The research script includes Newey–West statistics, Benjamini–Hochberg filtering, equal-weight and IC-weighted scores, ridge-based scores, and an optional time-based cross-validation path.

Reproduction prerequisites

The repository is not yet self-contained. Before running the pipeline, supply these input tables in the repository root:

Input	Required columns used by the code
constituents.csv	ticker, name
market_indices.csv	Country, ticker
stoxx_europe_600_v2.csv	Ticker_YF, Supersector, Country

These three tables are not committed. Ticker and country mappings must agree across them. The fetch script creates Prices.csv and index_data.csv.

requirements.txt records an existing environment, including the Windows-specific pywin32 package. Installation needs adaptation on other operating systems; a portable environment has not been verified.

Run sequence

After preparing the data inputs and Python environment, run from the repository root:

python 1_FETCH_DATA.py
python 2_DATA_ENG.py
python 3_EDA.py
python 4_ALPHA_RESEARCH.py

The default alpha input directory is 1_0_DATA_ENG; its output directory is 3_0_ALPHA_OUTPUTS. Inspect the available options with:

python 4_ALPHA_RESEARCH.py --help

The optional cross-validation path can be requested with:

python 4_ALPHA_RESEARCH.py \
    --indir 1_0_DATA_ENG \
    --outdir 3_0_ALPHA_OUTPUTS \
    --rebalance M \
    --topq 0.2 \
    --tc_bps 5 \
    --cv_enable \
    --cv_train_months 36 \
    --cv_test_months 6 \
    --cv_gap_months 1

The fetch script currently requests roughly three years of history. A 36-month training window plus gap and test period requires more usable history, especially after factor warm-up. Check whether valid folds were formed; the script logs a warning when none can be created.

Saved diagnostics

The committed IS_* charts are in-sample diagnostics. They are not verified out-of-sample performance. The script generates separate cv_* outputs when its optional cross-validation path produces results; those outputs are not present in the inspected repository.

Interpretation and limitations

• The run metadata describes a data-light setup without explicit delisting, corporate-action, earnings, or capitalization datasets. An observed-history universe is not a historical constituent database; survivorship effects remain possible.
• Return winsorization in add_effective_return uses quantiles from the supplied full panel before cross-validation. This requires revision before claiming fully independent out-of-sample evaluation.
• Returns prefer country-index excess returns where available and otherwise use raw returns. Interpret portfolio curves with that mixed return definition in mind.
• Fields labelled price_eur and ret_1d_eur do not by themselves establish currency conversion: the fetch script downloads Yahoo prices without an explicit FX step.
• The default cost setting is 5 basis points per unit turnover. It is a simplified research assumption.

A reproducible next iteration should supply documented universe mappings, freeze the data snapshot and environment, isolate preprocessing within training folds, and report benchmark-relative out-of-sample results with cost sensitivity.

License

See LICENSE.