Paper: arXiv 2408.10340

Authors: Nathalia Castellanos, Dhruv Desai, Sebastian Frank, Stefano Pasquali, Dhagash Mehta

Abstract

Peer analysis is a critical component of investment management, often relying on expert-provided categorization systems. These systems’ consistency is questioned when they do not align with cohorts from unsupervised clustering algorithms optimized for various metrics. We investigate whether unsupervised clustering can reproduce ground truth classes in a labeled dataset, showing that success depends on feature selection and the chosen distance metric. Using toy datasets and fund categorization as real-world examples we demonstrate that accurately reproducing ground truth classes is challenging. We also highlight the limitations of standard clustering evaluation metrics in identifying the optimal number of clusters relative to the ground truth classes. We then show that if appropriate features are available in the dataset, and a proper distance metric is known (e.g., using a supervised Random Forest-based distance metric learning method), then an unsupervised clustering can indeed reproduce the ground truth classes as distinct clusters.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced distance metric learning (Mahalanobis, RF-PHATE) and objective optimization formulas, indicating high mathematical density, but it lacks direct backtesting, live data pipelines, or portfolio implementation metrics, focusing instead on algorithmic reproduction on labeled datasets.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Can unsupervised clustering reproduce<br>expert-provided categorization?"] --> B

  subgraph B ["Methodology & Data"]
      direction LR
      B1["Toy Datasets<br>Controlled ground truth"] --> B2["Fund Categorization Data<br>Real-world labeled data"]
  end
  
  B --> C["Computational Process"]
  
  subgraph C ["Computational Process"]
      direction TB
      C1["Standard Clustering<br>Various distance metrics"] --> C2["Cluster Evaluation<br>Metric analysis vs ground truth"]
      C3["Supervised Learning<br>Random Forest<br>Distance Metric Learning"] --> C4["Optimized Clustering<br>Reproduction attempt"]
  end
  
  C --> D["Key Findings & Outcomes"]
  
  subgraph D ["Key Findings & Outcomes"]
      D1["Ground Truth Reproduction<br>Challenging with standard metrics"] --> D2["Success Condition<br>Requires appropriate features +<br>proper distance metric"]
  end

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 that with appropriate features and a learned distance metric (e.g., supervised Random Forest‑based), an unsupervised clustering algorithm can reproduce ground‑truth classes as distinct clusters.

Source sentence
we then show that if appropriate features are available in the dataset, and a proper distance metric is known (e.g., using a supervised Random Forest-based distance metric learning method), then an unsupervised clustering can indeed reproduce the ground truth classes as distinct clusters.

Assumes 4

  • The available variables are predictive enough for the ground‑truth labels. data
    Source sentence
    if the available variables are predictive enough for the ground truth labels and the distance metric is learned using the ground truth labels
  • A proper distance metric can be learned from the ground‑truth labels. model
    Source sentence
    the distance metric is learned using the ground truth labels with supervised distance metric learning methods such as the Mahalanobis distance metric learning
  • Standard K‑means clustering can be used as the base clustering algorithm. model
    Source sentence
    We first use Euclidean distance‑based clustering as our baseline, and then apply supervised distance metric learning with Mahalanobis distance
  • One‑hot encoding and zero imputation for missing categorical values are appropriate preprocessing steps. data
    Source sentence
    We one‑hot encoded all the categorical columns before employing any of the algorithms. There were no missing values in the toy datasets. For the fund’s data,

Data

  • Funds (Morningstar) · Morningstar fundamentals commercial
    March 2024 snapshot
    Source sentence
    imbalanced distribution, with 14 numerical and 2 categorical variables, for the March 2024 snapshot.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Standard Euclidean K‑means rarely reproduces the number of ground‑truth classes; internal metrics like Inertia and Davies‑Bouldin perform poorly. statistical
    Source sentence
    Inertia metric and Davies‑Bouldin index showed poor performance rarely reproducing the number of ground truth classes.
  • Using Mahalanobis distance improves internal metrics; Silhouette Score achieves exact match 53% of the time and discrepancy ≤2 in ~64% of datasets. statistical
    Source sentence
    Silhouette Score achieved an exact match 53% of the time and a discrepancy of 2 or less in around 64% of the datasets
  • RF‑PHATE transformed K‑means further improves matches; external metrics Fowlkes‑Mallows and Rand achieve exact matches around 75% of the time. statistical
    Source sentence
    the Fowlkes‑Mallows Score and Rand Score both achieved an exact match around 75% of the time
  • On the Morningstar fund dataset, advanced distance metrics achieve perfect match for six out of eight external evaluation metrics and are within one class for the remaining two. performance
    Source sentence
    obtaining the perfect match with the ground truth class count for six out of the eight external evaluation metrics, with the remaining two metrics being only one‑off from the ground truth class count

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive

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.