Paper: arXiv 2306.05568
Authors: Philippe Goulet Coulombe, Maximilian Goebel
Abstract
When it comes to stock returns, any form of predictability can bolster risk-adjusted profitability. We develop a collaborative machine learning algorithm that optimizes portfolio weights so that the resulting synthetic security is maximally predictable. Precisely, we introduce MACE, a multivariate extension of Alternating Conditional Expectations that achieves the aforementioned goal by wielding a Random Forest on one side of the equation, and a constrained Ridge Regression on the other. There are two key improvements with respect to Lo and MacKinlay’s original maximally predictable portfolio approach. First, it accommodates for any (nonlinear) forecasting algorithm and predictor set. Second, it handles large portfolios. We conduct exercises at the daily and monthly frequency and report significant increases in predictability and profitability using very little conditioning information. Interestingly, predictability is found in bad as well as good times, and MACE successfully navigates the debacle of 2022.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs only partially modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)
Why this score: The paper introduces MACE, a novel multivariate extension of Alternating Conditional Expectations, involving advanced concepts like constrained Ridge Regression and Random Forests, which requires dense mathematical formulation. It demonstrates strong empirical rigor through backtests across daily and monthly frequencies, using specific datasets (e.g., NASDAQ stocks) and reporting performance metrics like R² and Sharpe ratios, including results from the challenging 2022 period.
Research Flowchart
flowchart TD
A["Research Goal<br>Maximally Machine-Learnable Portfolios"] --> B["Key Methodology<br>MACE Algorithm"]
B --> C{"Process Flow"}
C --> D["Input: Asset Returns<br>& Predictors"]
C --> E["Random Forest<br>Nonlinear Feature Extraction"]
C --> F["Constrained Ridge Regression<br>Portfolio Optimization"]
D & E & F --> G["Output: Optimized Portfolio Weights"]
G --> H["Key Findings & Outcomes"]
H --> I["Predictability in<br>Good & Bad Times"]
H --> J["Significant Profitability<br>Boost"]
H --> K["Robust Performance<br>e.g., 2022 Debacle"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes MACE, a multivariate Alternating Conditional Expectations algorithm that jointly learns portfolio weights and a predictive function.
Source sentence
We introduce MACE, a multivariate extension of Alternating Conditional Expectations that achieves the aforementioned goal by wielding a Random Forest on one side of the equation, and a constrained Ridge Regression on the other.
Assumes 3
- An L2 regularization term guards against over‑fitting and unrealistic allocations. model
Source sentence
The addition of λ||w||2 provides l2 regularization with intensity λ (an hyperparameter) that will guard against overfitting and non‑realistic allocations
- Time‑series dependence in predictors can bias out‑of‑bag predictions, requiring careful handling. statistical
Source sentence
This persistence will break the non‑overfitting properties of OOB predictions – with an immediate consequence that f* s (Xt) includes overfitted elements
- Transaction costs are proportional to portfolio turnover with a spread‑based multiplier. execution
Source sentence
We derive TC as some multiple of portfolio turnover, where we estimate the multiple (c) as one‑half the bid‑ask spread
Data
- Welch and Goyal (2007) macro data set macro public
Source sentence
utilizes the canonical Welch and Goyal (2007) data set to construct an MMLP with a subset of large‑cap individual stock returns from CRSP
- CRSP large‑cap daily stock returns market-prices commercial
Source sentence
subset of large‑cap individual stock returns from CRSP
- Bid‑ask spreads of 100 stocks market-prices unknown
Source sentence
compute the distribution of c based on the bid‑ask spreads of all our 100 stocks, observed on a daily basis between 2017-01-01 and 2022-12-31
CRSP Kenneth French Data Library
Costs
Unclear - sources disagree
Source sentence (the authors' words)
take three candidates for c ∈{0.01%, 0.015%, 0.03%}, which correspond to the 50th, 75th, and 90th percentiles respectively
Finds 5
- MACE achieves an out‑of‑sample R² above 4% in the last 15 years of the monthly sample, outperforming other RF‑based portfolios. predictive-accuracy
Source sentence
MACE hits an R2 of above 4% in the last 15 years of our sample ending in 2019
- In the monthly exercise, MACE delivers an annualized average return of 18.70%, beating the next best model by six percentage points. performance
Source sentence
The annualized average return of rA = 18.70% beats the closest competitor (EW (RF)) by a stunning six percentage points
- MACE attains a Sharpe Ratio of 1.05, higher than its variants and benchmarks, while keeping drawdowns low. performance
Source sentence
both modified MACEs cannot keep up with the Sharpe Ratio of SR = 1.05
- Even with daily transaction costs up to 3 basis points, MACE remains highly competitive in absolute and risk‑adjusted returns. performance
Source sentence
Even though we are trading at a daily frequency, MACE remains highly competitive, both in absolute, as well as risk‑adjusted returns
- During the Covid‑19 recession, MACE generates massive gains linked to unusually high R² values. performance
Source sentence
MACEs generate massive gains during the short‑lived Pandemic recession. These can obviously be linked back to the abnormally high R2’s observed for that era
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.