Many of the models used in economic and financial analysis assume that their underlying parameters remain constant over time. But there’s growing evidence that these parameters often change, reflecting evolving market conditions, policy shifts or economic cycles.
Our research addresses the critical challenge of accurately estimating these time-varying parameters in complex, nonlinear time series models, without imposing rigid assumptions about how they change (Kristensen and Lee, 2025). By developing a novel theory, we aim to provide a more flexible and robust framework for understanding and predicting dynamic phenomena, ultimately offering a clearer picture of how economic relationships evolve.
What have we found out?
Our research introduces a sophisticated method for estimating parameters that change over time in a wide array of nonlinear time series models (including threshold autoregressions, ARCH models and Poisson autoregressions with exogenous covariates). The core of our approach lies in local polynomial estimation, which allows the parameters to vary smoothly over time without needing a pre-defined form for their variation.
A key finding of our theoretical analysis is the consistency and asymptotic normality (where a distribution gets closer to a normal/Gaussian as the sample size grows) of our proposed estimators. This means that they provide reliable estimates that follow a predictable statistical distribution in large samples, even under relatively weak conditions. Crucially, we offer the first precise characterisation of the leading bias term resulting from the smoothing process. This has not been achieved before and is vital for accurate inference.
Our theory also reveals that local linear estimators generally outperform local constant estimators. Specifically, local linear estimators require weaker regularity conditions, they suffer from fewer biases in the interior of the data range and they exhibit an automatic boundary adjustment property. This last feature is particularly valuable as it improves estimation accuracy at the beginning and end of the observed data – which are often points of significant interest.
Our methods and theoretical results also apply to discrete-valued time series models (such as Poisson autoregressions), which existing theories struggle to accommodate due to their reliance on strong smoothness conditions. Our new proof techniques enable the application of local linear estimators to these models without modification.
Why is this important?
Our research significantly advances the field of econometrics and time series analysis. By providing a robust and flexible framework for estimating time-varying parameters, we move beyond the limitations of traditional parametric models that assume constant parameters or impose specific forms of time variation. This is crucial because mis-specified models can lead to inaccurate forecasts and flawed policy recommendations.
Our ability to characterise precisely the leading bias term enables more accurate statistical inference and the use of standard bandwidth selection rules, which were previously problematic in this context. The applicability to discrete-valued time series models opens up new avenues for analysing a broader range of real-world phenomena, such as corporate default counts (which are inherently discrete).
Indeed, an empirical application of our approach to US corporate default counts vividly illustrates the practical importance of our findings. We use a time-varying Poisson autoregression model and find substantial evidence of time variation in the parameters that previous studies, assuming constant parameters, could not capture.
For example, the influence of macroeconomic and financial variables on corporate defaults has not been static but has changed considerably over time. This dynamic understanding can inform financial risk management, economic policy and regulatory frameworks by providing a more nuanced view of market behaviour and interconnectedness. As our analysis suggests, neglecting these time variations can lead to an overestimation of financial contagion.
What next?
Based on our findings, there are several paths for future research and policy considerations.
From a research perspective, while we’ve established primitive conditions for Markov models with exogenous covariates (models where how likely a system is to move to the next state depends on external factors as well as past states), further work is needed to establish conditions under which other complex time series, particularly infinite memory processes, satisfy our definition of local stationarity (where statistical properties like the mean, variance and autocorrelation change slowly over time, allowing the series to appear stationary within small, localised windows but non-stationary globally).
In addition, exploring the application of local polynomial estimators to a wider range of nonlinear models and developing formal statistical tests for the significance of time-varying parameters would be beneficial. Investigating higher-order local polynomial estimators and their practical implications could also refine the methodology.
For policy and practical applications, the demonstrated time variation in economic relationships highlights the need for adaptive modelling approaches in forecasting and policy design. Policy-makers should be aware that the impact of economic variables can change over time, and models used for decision-making should ideally incorporate this dynamism.
For financial institutions, incorporating time-varying parameter models could lead to more accurate risk assessments and capital allocation strategies, especially during periods of market volatility. Regulatory bodies could also make use of these more sophisticated models to develop more responsive and effective systems that account for evolving market dynamics rather than relying on static assumptions.
