Paper: arXiv 2501.00034
Authors: Chengyue Huang, Yahe Yang
Abstract
With the widespread application of machine learning in financial risk management, conventional wisdom suggests that longer training periods and more feature variables contribute to improved model performance. This paper, focusing on mortgage default prediction, empirically discovers a phenomenon that contradicts traditional knowledge: in time series prediction, increased training data timespan and additional non-critical features actually lead to significant deterioration in prediction effectiveness. Using Fannie Mae’s mortgage data, the study compares predictive performance across different time window lengths (2012-2022) and feature combinations, revealing that shorter time windows (such as single-year periods) paired with carefully selected key features yield superior prediction results. The experimental results indicate that extended time spans may introduce noise from historical data and outdated market patterns, while excessive non-critical features interfere with the model’s learning of core default factors. This research not only challenges the traditional “more is better” approach in data modeling but also provides new insights and practical guidance for feature selection and time window optimization in financial risk prediction.
Complexity vs Empirical Score
- Math Complexity: 4.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper uses standard machine learning models (Logistic Regression, Random Forest, LSTM, Transformer) and focuses on empirical analysis of feature sets and time windows, with no advanced derivations. However, it demonstrates high empirical rigor through extensive backtesting on a large real-world dataset (6.6 million observations from Fannie Mae), controlled experiments, and clear practical guidelines.
Research Flowchart
flowchart TD A["Research Goal: Investigate the Impact of<br>Training Timespan & Feature Count on<br>Mortgage Default Prediction Performance"] --> B["Data Source & Period<br>Fannie Mae Mortgage Data (2012-2022)"] B --> C["Key Methodology: Controlled Experiments<br>Comparing Short vs. Long Time Windows"] C --> D["Computational Process:<br>Model Training on Varied Feature Sets"] D --> E["Findings: Shorter Time Windows<br>& Critical Features Yield<br>Superior Prediction Performance"] E --> F["Outcome: Contradicts 'More is Better'<br>Empirical Solution for Feature/Time<br>Window Optimization"]
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 empirically validates a “feature redundancy paradox” and offers practical guidelines for optimizing feature and time‑window selection in mortgage default prediction.
Source sentence
By synthesizing insights from prior studies, our work aims to empirically validate the ”feature redundancy paradox” and provide actionable guidelines for optimizing feature and time window selection in mortgage default prediction.
Assumes 3
- Housing markets are dynamic and evolving, so older patterns may become irrelevant. market
Source sentence
A key difficulty lies in the dynamic and evolving nature of housing markets and broader economic conditions.
- Including many non‑critical features can obscure the influence of core variables. model
Source sentence
Likewise, adding numerous non‑essential features can obscure the influence of critical variables and hinder the model’s ability to i…
- Shorter, more recent time windows provide temporally relevant information that improves forecasting accuracy. data
Source sentence
Studies have shown that incorporating too much historical data can introduce noise and outdated patterns, which may degrade model performance.
Data
- Freddie Mac single‑family mortgage performance dataset other unknown
Source sentence
Our study utilizes Freddie Mac’s single‑family mortgage monthly performance data spanning from 2012 to 2022, comprising approximately 6.6 million monthly observations from 550,000 mortgage loans (50,000 loans sampled per year).
Costs
Not applicable
Finds 5
- Logistic Regression’s ROC‑AUC drops from 0.858 to 0.803 when moving from a one‑year to a ten‑year training window. performance
Source sentence
Logistic Regression’s performance drops from 0.858 to 0.803, and Random Forest from 0.865 to 0.805 when moving from one-year to ten‑year windows.
- The key feature subset (10 features) consistently outperforms the full (26) and reduced (18) feature sets, improving ROC‑AUC by 0.016 to 0.029 in short windows. predictive-accuracy
Source sentence
The key feature subset (n=10) consistently outperforms both the full feature set (n=26) and the reduced feature set (n=18), with improvements ranging from 0.016 to 0.029 in short windows.
- Transformer models retain higher ROC‑AUC across longer windows compared to LSTM and other deep‑learning models. performance
Source sentence
Transformer models demonstrate greater robustness to variations in time window lengths compared to LSTM and Transformer models in time series forecasting.
- Across all architectures, feature reduction yields beneficial results, with the reduced set (18) serving as an effective intermediate step. predictive-accuracy
Source sentence
The reduced feature set (n=18) often serves as an effective intermediate step, suggesting a gradual benefit to feature parsimony.
- Longer historical periods introduce temporal noise that outweighs the benefit of additional training samples. theoretical
Source sentence
while longer historical data periods may provide additional training samples, the potential benefits are outweighed by the introduction of temporal noise and outdated patterns.
Checked by others
Supported by
- Conflicting History in Foundation Model Return Forecasts
Compare the findings
This paper: Longer historical periods introduce temporal noise that outweighs the benefit of additional training samples.
That paper: Conflicting older history reduces the probability of improving on zero in all 30 higher‑signal comparisons across five models and three processes.
1 paper 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.