Paper: arXiv 2308.15384
Authors: Elliot Beck, Damian Kozbur, Michael Wolf
Abstract
This papers proposes a generic, high-level methodology for generating forecast combinations that would deliver the optimal linearly combined forecast in terms of the mean-squared forecast error if one had access to two population quantities: the mean vector and the covariance matrix of the vector of individual forecast errors. We point out that this problem is identical to a mean-variance portfolio construction problem, in which portfolio weights correspond to forecast combination weights. We allow negative forecast weights and interpret such weights as hedging over and under estimation risks across estimators. This interpretation follows directly as an implication of the portfolio analogy. We demonstrate our method’s improved out-of-sample performance relative to standard methods in combining tree forecasts to form weighted random forests in 14 data sets.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical concepts like mean-variance optimization and convex optimization with L1 constraints, demonstrating high mathematical density. It also validates the method on 14 benchmark datasets with out-of-sample performance metrics, showing solid empirical rigor though not as extensive as full-scale trading backtests.
Research Flowchart
flowchart TD A["Research Goal:<br>Optimal Forecast Combination"] --> B["Identify Problem:<br>Mean-Variance Portfolio Analogy"] B --> C["Methodology:<br>Estimate Population Mean & Covariance"] C --> D["Computational Process:<br>Compute Optimal Hedge Weights"] D --> E["Data Input:<br>14 Datasets for Weighted Random Forests"] E --> F["Application:<br>Combine Tree Forecasts"] F --> G["Outcome:<br>Improved Out-of-Sample Performance"] G --> H["Key Finding:<br>Hedging Weights Outperform Standard Methods"]
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 a generic, high‑level methodology for generating forecast combinations that achieve the optimal linear combination in terms of mean‑squared forecast error, using the mean vector and covariance matrix of individual forecast errors and allowing negative weights.
Source sentence
This paper proposes a generic, high-level methodology for generating forecast combinations that would deliver the optimal linearly combined forecast in terms of the mean-squared forecast error if one had access to two population quantities: the mean vector and the covariance matrix of the vector of individual forecast
Assumes 6
- The observations {v_i} are an independent and identically distributed (i.i.d.) sample. data
Source sentence
is an independent and identically distributed (i.i.d.) sample where the distribution
- The mean vector µ and covariance matrix Σ of forecast errors can be estimated from training data. statistical
Source sentence
In practice, the quantities needed above are unknown and must be estimated based on a set of available (training) data.
- Shrinkage estimation is used for the covariance matrix instead of the sample covariance matrix. model
Source sentence
One of our contributions is to suggest the use of nonlinear shrinkage in order to estimate the covariance matrix as opposed to the standard (or canonical) choice in the related literature, the sample covariance matrix.
- Negative forecast combination weights are allowed. model
Source sentence
We allow negative forecast weights and interpret such weights as hedging over and under estimation risks across estimators.
Show 2 more
- Training‑set sizes are bounded between 200 and 5,000 observations. data
Source sentence
Throughout, we consider a fixed range of training-set (sample) sizes 200 ≤ n ≤ 5,000.
- The random forest uses 500 trees with default hyper‑parameters from the ranger library. execution
Source sentence
the number of trees is set to p = 500 as per default.
Data
- 14 PMLB benchmark data sets benchmark public
Source sentence
we use 14 data sets sourced from the Penn Machine Learning Benchmarks (PMLB) database; see Romano et al. (2021).
Costs
Not applicable
Finds 4
- The hedged random forest (HRF) clearly outperforms the standard random forest (RF) on balance across the 14 data sets. performance
Source sentence
On balance, HRF clearly outperforms RF.
- Performance gains are most pronounced for small training‑set sizes but persist up to n = 5,000. performance
Source sentence
The gains are most pronounced for small training-set sizes n but ‘live on’ up to the largest size considered, n = 5000.
- HRF performs worse than RF on only one data set, with a loss never exceeding 6% (and always below 5% for n ≥ 400). negative
Source sentence
Out of the 14 data sets, there is one on which HRF performs worse than RF, but the loss is never more than 6% and always below 5% for n ≥400.
- For two data sets HRF reduces RMSE by more than 25% for n ≥ 400, and for one of them the reduction exceeds 40% for all n. performance
Source sentence
there are two data sets for which HRF reduces the RMSE compared to RF by more than 25% for n ≥400; for one of these data sets, HRF actually reduces the RMSE by more than 40% for all n.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 19 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.