Paper: arXiv 2405.03402

Authors: Etienne Theising

Abstract

This paper introduces an approach to reference class selection in distributional forecasting with an application to corporate sales growth rates using several co-variates as reference variables, that are implicit predictors. The method can be used to detect expert or model-based forecasts exposed to (behavioral) bias or to forecast distributions with reference classes. These are sets of similar entities, here firms, and rank based algorithms for their selection are proposed, including an optional preprocessing data dimension reduction via principal components analysis. Forecasts are optimal if they match the underlying distribution as closely as possible. Probability integral transform values rank the forecast capability of different reference variable sets and algorithms in a backtest on a data set of 21,808 US firms over the time period 1950 - 2019. In particular, algorithms on dimension reduced variables perform well using contemporaneous balance sheet and financial market parameters along with past sales growth rates and past operating margins changes. Comparisions of actual analysts’ estimates to distributional forecasts and of historic distributional forecasts to realized sales growth illustrate the practical use of the method.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced statistical concepts like distributional forecasting, principal component analysis, and probability integral transform values for evaluation, indicating moderate-to-high math complexity. It demonstrates high empirical rigor by backtesting on a massive dataset (21,808 US firms, 1950-2019) and comparing results against actual analyst estimates.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Distributional Forecasting of<br>Corporate Sales Growth"] --> B["Data & Inputs<br>21,808 US Firms<br>1950-2019<br>Co-variates: Balance Sheet,<br>Market Parameters, History"]
  B --> C{"Methodology"}
  C --> D["Dimension Reduction<br>Principal Component Analysis<br>Optional Step"]
  C --> E["Reference Class Selection<br>Rank-based Algorithms"]
  D --> F
  E --> F["Distributional Forecasts<br>Ranking via Probability<br>Integral Transform PIT"]
  F --> G["Backtesting & Evaluation<br>Compare Forecasts vs.<br>Actual Realized Growth"]
  G --> H["Key Findings/Outcomes<br>Dimension Reduction +<br>Contemporaneous Data performs best.<br>Effective for bias detection &<br>analyst comparison"]

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 rank‑based reference class selection methods that use multiple reference variables and optionally apply principal component analysis for dimension reduction.

Source sentence
The novel rank based methods allow several reference variables and include the option of a dimension reduction based on principal components.

Assumes 3

  • A smooth data‑generating function f_h exists such that future sales growth conditional on reference variables follows this function. model
    Source sentence
    The first assumption behind the approach is the existence of a data generating mechanism, say a smooth function fh such that Yi,t+h ∼fh(Xi,τ:t) which can be interpreted as the conditional distribution Yi,t+h|Xi,τ:t.
  • The data‑generating mechanism is stable over time (stationarity) so that the same function applies to reference class observations. statistical
    Source sentence
    we need an assumption that this mechanism works similarly over time and we have Yj,s+h ∼fh(Xj,ζ:s), (j, s) ∈R, for the outcomes within the reference class, which can be interpreted as a kind of stationarity assumption.
  • Sufficient historical firm data are available to assess the distribution of future sales growth. data
    Source sentence
    To this end, we assume that a sufficient amount of historical data on firms is available to assess the distribution of Yi,t+h.

Data

  • Compustat North America fundamentals · S&P Global Market Intelligence fundamentals commercial
    1950-2019 · US firms excluding financial and real‑estate sectors
    Source sentence
    This data set consists of Compustat North America fundamentals annual data from 1950 to 2019 by S&P Global Market Intelligence (2020) limited to US firms and excluding companies from the financial and real‑estate sector.
  • CRSP daily stock file · Center for Research in Security Prices (CRSP) market-prices commercial
    1950-2019 · US firms (21,808) with stock exchange information
    Source sentence
    These data are merged with stock‑exchange information from the Center for Research in Security Prices (CRSP, 2020) daily stock of the University of Chicago Booth School of Business.

Recognised sources: WRDS FactSet Compustat

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • For a one‑year forecast horizon, the PCARD method improves the ∆q metric by 71% while using five years less data. predictive-accuracy
    Source sentence
    For one-year forecast horizon (see Table 6) PCARD improves ∆q by 71% and RD improves ∆q by 58% while using five years less and one year more of data, respectively.
  • Principal component analysis reduces the amount of necessary past data and improves forecast performance by 38% to 71% across horizons. predictive-accuracy
    Source sentence
    principal component analysis reduces the amount of necessary past data while simultaneously improving previous approaches (Theising et al., 2023) substantially by between 38% and 71% depending on forecast horizon.
  • For three‑year ahead forecasts, PCARD reduces ∆q by 50% with 14 years less data, and RD reduces ∆q by 23% with three years more data. predictive-accuracy
    Source sentence
    When forecasting three years ahead, Table 7 shows a reduction of ∆q by 50% and by 23% for PCARD and RD with 14 years less and three years more of data, respectively.
  • For ten‑year forecasts, contemporaneous balance‑sheet variables with PCARD reduce ∆q by 38% using similar data, while RD improves ∆q by 12% with comparable data demand. predictive-accuracy
    Source sentence
    In Table 9 on 10-year forecast horizon it stands out that contemporaneous balance sheet variables and past variables are the best reference variables for PCARD and reduce ∆q by 38% with similar necessary data. While the best RD combinations have a comparable data demand as the single variable approach and improve ∆q by

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 16 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.