US Firm Characteristics
Classic factor investing with ML on monthly fundamental data
Fundamentals
Monthly
Fundamental Data
Methodology Highlight
Demonstrates how label engineering, preprocessing choices, and point-in-time data management affect results as much as model selection — with the strongest statistical foundation (10 CV folds) in the book.
This case study applies ML to the canonical factor investing question: can machine learning improve on traditional long-short decile sorts when accounting lags, survivorship bias, and transaction costs are taken seriously? Working with 57 firm-level characteristics spanning valuation, profitability, momentum, and risk across approximately 2,500 US stocks (1996-2016), this is the most feature-rich fundamental dataset in the book.
Students learn point-in-time data management with accounting publication lags, the impact of label engineering choices (classification vs regression), and the effect of preprocessing decisions like winsorization on model performance. The case study provides the strongest statistical foundation with 10 CV folds, enabling students to assess the reliability of walk-forward results.
The fundamental universe is the natural home for latent factor models (IPCA, CAE, SAE, SDF). Students learn how these architectures extract time-varying factor loadings from cross-sectional data, and how capacity constraints in small-cap stocks limit the practical scalability of strategies that appear strong on paper.
Strategy Summary
Long-short decile sort on approximately 2,500 US stocks using ML predictions on winsorized forward returns. Monthly rebalancing with 6-month characteristic lag enforced for point-in-time compliance. Equal-weight within deciles, dollar-neutral. 10 CV folds with 10-year training and 1-year validation windows provide the most robust statistical evaluation in the book.
Data Sources
NASDAQ Data Link (firm characteristics)
ML Techniques
GBM classification
Latent factor models (IPCA, CAE, SAE, SDF)
Label engineering (classification vs regression)
Cross-sectional prediction
ML Pipeline
Universe & Setup
~2,500 US stocks filtered by price >$5 and ADV >$1M. Monthly decision cadence
with 6-month accounting lag enforced for point-in-time compliance. The longest
history (1996-2016) and most CV folds (10, with 10Y train / 1Y val) of any case
study. Era-dependent costs: pre-decimalization spreads 15-30 bps, post-2001
spreads 5-15 bps. Anonymized NASDAQ Data Link characteristics.
Universe & Protocol Setup
Ch 6
Defines the trading universe (US stocks filtered by price >$5 and ADV >$1M),
monthly decision cadence, and 6-month accounting lag convention. Validates
feasibility against transaction costs and builds walk-forward evaluation
splits with purging. Documents the anonymized PERMNO-based identifier system.
Labels & Evaluation
1-month forward return as primary label with a winsorized classification variant
that produces the largest label engineering effect in the book: IC jumps ninefold
from 0.008 to 0.074. 3-month variant for horizon sensitivity. No temporal features
(no Ch9 notebook) -- the 57 firm characteristics are rich enough. Evaluation
validates classification dominance across all feature families.
Label Engineering
Ch 7
Computes 1-month forward returns and classification labels from the Chen-
Pelger-Zhu dataset where characteristics at month t-1 are paired with returns
during month t. Creates winsorized and median-split classification variants.
The winsorized classification label (fwd_class_1m) produces a dramatic IC
improvement over raw returns -- the largest label engineering effect in the
book.
Feature Evaluation
Ch 7
Evaluates financial features with HAC-adjusted IC against 1-month forward
returns. Applies Benjamini-Hochberg FDR to control false discovery across the
factor zoo. Assesses quantile monotonicity and fold-level stability across
walk-forward folds. Produces triage decisions (PROCEED/REVISE/STOP) for
Ch11 model selection.
Feature Engineering
57 pre-computed Chen-Pelger-Zhu firm characteristics spanning five families:
fundamentals (book-to-market, earnings yield, cash flow), profitability (ROE, ROA,
gross margin), momentum (12-1 month, 6-month, reversal), risk (beta, idiosyncratic
vol, leverage), and investment (asset growth, R&D, CapEx). No temporal feature
notebook -- the feature set is entirely cross-sectional.
Feature Engineering
Ch 8
Maps pre-computed Chen-Pelger-Zhu firm characteristics into five economic
factor families: value (BEME, E2P), quality/profitability (ROE, OP),
investment (NOA, DPI2A), momentum (r12_2, ST_REV), and risk/liquidity (Beta,
IdioVol). Engineers composite factor scores (value+quality, value+momentum)
and interaction features (value x quality, momentum x ivol) for downstream
modeling.
Modeling
The natural home for latent factor models with 5 dedicated notebooks: standard
latent factors, IPCA (time-varying loadings), CAE (non-linear generalization of
IPCA), SAE (return-targeted extraction), and SDF (pricing kernel from
characteristics). GBM achieves IC +0.070 on classification labels. Classification
dominates regression everywhere. Causal DML tests whether 12-1 momentum causes
returns or reflects confounders.
Linear Models
Ch 11
Trains Ridge, LASSO, and ElasticNet on the firm characteristics panel with
walk-forward CV (10-year training windows). Reveals that classification labels
dominate regression: winsorized classification produces dramatically higher IC
than raw returns. Establishes the linear baseline that GBM and latent factor
models must beat.
Gradient Boosting
Ch 12
Trains LightGBM across leaf-count profiles and loss functions on fundamental
and price features. Searches for value-quality-momentum interactions
following Freyberger, Neuhierl, and Weber (2020). GBM on winsorized returns
substantially outperforms the ridge baseline. The IC gap between raw and
winsorized labels reinforces that label treatment is more impactful than
model architecture.
Tabular Deep Learning (TabM)
Ch 12
Trains TabM rank-1 adapter MLP ensemble (small/medium/large) on the flat
characteristic feature matrix as a neural alternative to GBM. Compares
attention-based tabular DL against tree-based models where GBM already leads
on both winsorized returns and classification labels.
Latent Factor Models
Ch 14
Runs PCA, IPCA, CAE, SDF, and SAE latent factor extraction on the richest
characteristic panel in the book. Compares
pricing-theory-motivated (SDF) vs statistical (CAE/SAE) factor extraction and
evaluates whether dimensionality reduction improves monthly return prediction
over direct GBM/TabM.
IPCA
Ch 14
Fits Instrumented PCA (Kelly, Pruitt, and Su 2019) with time-varying factor
loadings as linear functions of observable characteristics. The natural
model for this dataset: anonymized stock_id structure prevents standard PCA
from tracking entities across months, making IPCA's characteristic-conditional
approach essential. Estimates and interprets the Gamma matrix.
Conditional Autoencoder (CAE)
Ch 14
Trains the Conditional Autoencoder (Gu, Kelly, and Xiu 2021), replacing
IPCA's linear Gamma mapping with a neural network encoder for
characteristic-to-factor-loading estimation. Evaluates whether non-linear
interactions between characteristics (value-quality-momentum) improve
prediction over the linear IPCA baseline on identical folds.
Supervised Autoencoder (SAE)
Ch 14
Trains the Supervised Autoencoder combining reconstruction loss (preserve
characteristic structure) with prediction loss (forecast returns) in a
multi-task framework. The reconstruction term regularizes latent
representations against overfitting to noisy monthly returns. Completes the
autoencoder comparison before SDF provides the pricing-theory perspective.
SDF Network
Ch 14
Trains the Stochastic Discount Factor network to learn portfolio weights
satisfying the no-arbitrage condition E[M_t * R_{i,t}] = 0. Identifies which
firms are exposed to systematic risk -- a pricing-theory complement to the
supervised prediction objective used by IPCA, CAE, and SAE.
Causal DML
Ch 15
Applies DML to estimate the causal effect of 12-month momentum (r12_2) on
monthly returns, conditioning on Fama-French confounders (Beta, IdioVol, LME,
Variance). Finds a significant causal effect with partial confounding
absorption -- supporting the behavioral explanation over the risk-premium
explanation.
Strategy Pipeline
Long-short decile strategy simulation. Holdout Sharpe +2.52 (17% decay -- mildest
in book). Risk overlay lifts managed Sharpe to +3.92. Era-dependent cost analysis
spanning pre- and post-decimalization. Critical capacity finding: 69% of long leg
in bottom market-cap quartile where spreads run 100-500 bps.
Model Analysis
Ch 11
Compares all model families on the classic factor zoo dataset. TabM leads on
raw returns with the only compelling standalone decile spread. GBM on
winsorized returns outperforms all family-label combinations. Documents that
label design dominates model choice. Four families advance (TabM, GBM, SDF,
causal DML), one excluded (linear).
Backtest & Signal Evaluation
Ch 16
Runs plumbing test, parametric sweep across all prediction-signal
combinations, and statistical analysis (DSR, family comparison) for a long-
short decile strategy on US stocks. Translates the classification signal
into simulated portfolio returns across the full backtest sweep with
walk-forward CV.
Portfolio: Allocator Sweep
Ch 17
Sweeps top predictions x TOP_K concentration x 6 allocators (equal-weight
through HRP) on the stock universe. Evaluates concentration vs
diversification tradeoffs: holding too few stocks sacrifices diversification,
while holding too many dilutes signal. Compares risk-parity and score-weighted
allocation against equal-weight.
Transaction Costs
Ch 18
Sweeps the cost grid on top allocation combinations, confirming the most
favorable cost profile in the book: monthly rebalancing provides much lower
turnover than daily strategies with ample headroom. Analyzes era-dependent
costs spanning pre- and post-decimalization.
Risk Management
Ch 19
Sweeps position-level (stop-loss, trailing stops) and portfolio-level
(drawdown breakers, daily loss limits) risk controls on the strongest
strategy in the book. Evaluates whether risk overlays add robustness to an
already stable monthly-cadence strategy with modest holdout decay.
Synthesis & Verdict
Strongest integrated result in the book. Verdict: Advance -- but filter universe
to top three market-cap quartiles and re-evaluate under realistic capacity
constraints. Factor attribution confirms genuine alpha (t=7.09).
Strategy Analysis
Ch 20
Synthesizes the end-to-end results for the book's strongest case study. GBM
champion on fwd_class_1m produces strong holdout Sharpe before capacity
constraints. Produces the "advance" verdict with a capacity caveat: the long
leg concentrates in small-cap stocks, requiring universe filtering for
realistic deployment.
Quick Info
Asset Class
Fundamentals
Frequency
Monthly
Data Type
Fundamental Data
Notebooks
19
Chapters
12
Chapters
6
Strategy Research Framework
7
Defining the Learning Task
8
Financial Feature Engineering
11
The ML Pipeline
12
Advanced Models for Tabular Data
14
Latent Factor Models
15
Causal Machine Learning
16
Strategy Simulation
17
Portfolio Construction
18
Transaction Costs
19
Risk Management
20
Strategy Synthesis