TL;DR
This paper introduces a novel constrained regression framework for interval-valued data, specifically addressing the issue of maintaining the natural order (lower bound < upper bound) in regression models. They propose a two-step estimation procedure that is consistent and more robust than existing methods, particularly in high-volatility financial data.
What Problem Does This Paper Address?
Many real-world datasets, especially in finance, economics, and environmental science, are naturally represented as intervals (e.g., daily price ranges, temperature intervals). Standard regression techniques applied to these interval-valued data often fail to guarantee a fundamental property: the predicted lower bound of an interval is always less than or equal to the predicted upper bound. This violation can lead to nonsensical results. The authors tackle this by developing a constrained regression model that explicitly enforces this natural ordering, especially crucial for interval time series where forecasts must also respect this property.
Methods
The core of the paper lies in proposing a constrained regression model for interval time series. The key innovation is to model the upper and lower bounds of the intervals jointly, recognizing that the constraint $y_{l,t} \le y_{u,t}$ leads to a conditionally truncated bivariate normal distribution for the error terms. This truncation makes standard Ordinary Least Squares (OLS) estimators inconsistent.
To overcome this, they propose two estimation strategies:
- Two-Step Estimation (TS): This method first estimates the severity of the truncation (using maximum likelihood on the range of the interval) and then uses this information to correct for selection bias in a second step, typically involving least squares.
- Modified Two-Step Estimation (MTS): This is an extension of the TS method designed to handle multicollinearity issues that can arise in the second step, especially when the truncation is severe. It employs a minimum distance estimator in the second step, which is more robust to identification problems.
Datasets
The authors illustrate their methodology using the daily low/high returns of the SP500 index from January 1, 2004, to April 29, 2011. They split this data into two periods:
- Stable Period (2004-2007): Characterized by low volatility.
- Unstable Period (2007-2011): Characterized by high volatility, including the 2008 financial crisis.
They also conducted extensive Monte Carlo simulations using various data generating processes (DGPs) to evaluate the performance of their proposed methods against existing ones.
Key Insights and Ideas
- The "Natural Order" Matters: Ignoring the inherent ordering of interval bounds ($y_l \le y_u$) in regression models leads to inconsistent estimators, especially when the data exhibits significant volatility or when the constraint is binding.
- Truncation is Key: The constraint $y_l \le y_u$ induces a truncated bivariate distribution for the model's errors, fundamentally altering their properties and invalidating standard OLS.
- Two-Step is Better than OLS: While OLS might appear adequate when the constraint is not binding (low volatility), it fails dramatically when the constraint is binding (high volatility). The proposed two-step methods provide consistent estimates in both scenarios.
- Modified Two-Step for Robustness: The MTS approach is crucial when multicollinearity arises in the second step of the estimation, ensuring reliable parameter identification even under severe truncation.
- Financial Crises Expose Weaknesses: The SP500 data analysis clearly shows that the 2008 financial crisis period (high volatility) severely violates the assumptions of standard OLS, highlighting the need for constrained methods.
Key Results
- Consistency: Both the two-step (TS) and modified two-step (MTS) estimators are shown to be consistent.
- Superior Performance: Monte Carlo simulations and the SP500 return analysis demonstrate that TS and MTS consistently outperform OLS and other existing interval regression methods (like CCRM, CRM, and GARCH-based models) in terms of lower losses (RMSE, MLF, MDE) and better estimation accuracy (lower bias, lower MSE).
- SP500 Analysis: In the unstable period (2007-2011), the estimated inverse Mill's ratio ($\lambda_{t-1}$) was significantly non-zero, indicating severe truncation. OLS estimates were found to be unreliable, while MTS provided robust and consistent estimates. In the stable period, where truncation was minimal, OLS was acceptable, but MTS still performed comparably.
Limitations and Future Work
- Computational Burden: While easier to implement than full maximum likelihood, the two-step procedures can still be computationally intensive, especially the MTS.
- Normality Assumption: The core methodology relies on a bivariate normal distribution assumption for the errors, although simulations suggest some robustness to misspecification.
- Future Work: The authors suggest exploring extensions to non-linear models and investigating the performance of their methods with different distributional assumptions for the errors. They also highlight the need for further research into handling multicollinearity more efficiently.
Summary updated