A Projection-Based ARIMA Framework for Nonlinear Dynamics in Macroeconomic and Financial Time Series: Closed-Form Estimation and Rolling-Window Inference

This paper introduces Galerkin-ARIMA and Galerkin-SARIMA, a projection-based framework that enhances classical time series models with low-dimensional basis expansions to better capture nonlinear dynamics while maintaining closed-form estimation, asymptotic consistency, and efficient rolling-window inference for macroeconomic and financial forecasting.

The Gist

Imagine you are trying to predict the weather. For decades, meteorologists have used a very specific, rigid rulebook: "If it rained yesterday, it will rain today with a slight adjustment." This rulebook is called ARIMA (Autoregressive Integrated Moving Average). It's the workhorse of economics and finance, used by central banks to set interest rates and by traders to guess stock prices.

The problem? The real world isn't always a straight line. Sometimes, the economy doesn't just "drift"; it jumps, loops, or reacts differently depending on how hot or cold things are getting. The old rulebook is too stiff to capture these twists and turns.

Enter Galerkin-ARIMA, a new method proposed by Liu and Lin. Think of it as taking that rigid rulebook and upgrading it with a Lego set.

The Core Idea: From a Ruler to a Flexible Ruler

1. The Old Way (Classical ARIMA): The Rigid Ruler
Imagine trying to draw a curvy mountain range using only a straight ruler. You can get close, but you'll always have jagged gaps. Classical ARIMA tries to predict the future by drawing a straight line through past data points. It's fast and easy to understand, but if the data curves, the prediction misses the mark.

2. The New Way (Galerkin-ARIMA): The Flexible Ruler (The "Lego" Approach)
The authors introduce a method called Galerkin projection. Imagine instead of a single straight ruler, you have a box of Lego bricks (or a flexible measuring tape made of many small segments).

• You can snap these bricks together to form a straight line if the data is simple.
• But if the data curves, you can bend the Lego structure to fit the curve perfectly.
• Crucially, you can still snap it apart and reassemble it instantly.

In math terms, instead of assuming the future is just a straight sum of the past, this method assumes the future is a smooth, flexible shape built from a few basic building blocks (polynomials or splines). It captures the "curves" in the economy that the old model misses.

The Secret Sauce: The Two-Stage "Assembly Line"

Here is the genius part. Usually, when you make a flexible model like this, it becomes a nightmare to calculate. You have to solve complex, messy equations every time you get a new piece of data. This is why most flexible models are too slow for real-time trading or daily central bank updates.

The authors solved this with a Two-Stage Assembly Line:

• Stage 1 (The Rough Draft): They use the Lego bricks to guess the general shape of the trend. This is a simple, straight-line calculation (like solving a basic math problem).
• Stage 2 (The Fine-Tuning): They look at the mistakes (residuals) from the first guess and use another set of Lego bricks to fix those specific errors.

Why is this cool?
Because both stages are just simple linear algebra. It's like solving a Sudoku puzzle where the rules never change, rather than trying to solve a Rubik's cube that changes its rules every second.

• Result: They can update their prediction thousands of times faster than the old method. While the old method might take a minute to re-calculate a forecast, the new method does it in a fraction of a second.

Real-World Analogies

• The Stock Market: Imagine a trader trying to predict the S&P 500. The market is noisy and sometimes behaves weirdly. The old model is like a driver who only knows how to drive in a straight line. The Galerkin model is like a driver with a GPS that can instantly recalculate a route around a sudden traffic jam or a pothole, without needing to stop the car to reprogram the engine.
• The Central Bank: Imagine the Federal Reserve needs to predict inflation every week. They have a huge window of data that slides forward every day. The old method is like a chef who has to chop all the vegetables from scratch every single morning. The new method is like a chef who has a pre-chopped, pre-organized kit that just needs a quick stir. It's the same meal, but it's ready in seconds.

What Did They Find?

The authors tested this on fake data (where they knew the answer) and real data (US GDP, Unemployment, and S&P 500).

1. Accuracy: When the economy was behaving "normally" (straight lines), the new method was just as good as the old one. But when the economy got "weird" (non-linear, bumpy), the new method was significantly more accurate because it could bend to fit the data.
2. Speed: The new method was orders of magnitude faster. It could update forecasts almost instantly, making it perfect for high-speed trading or daily policy decisions.
3. Stability: Sometimes, using too many "Lego bricks" makes the model wobbly. The authors added a "Ridge Regularization" feature, which is like adding shock absorbers to the car. It keeps the model stable even when the data is messy, preventing wild, unrealistic predictions.

The Bottom Line

Galerkin-ARIMA is a bridge between the old, reliable, but stiff world of classical statistics and the new, flexible world of machine learning.

• It keeps the interpretability of the old models (you can still understand why it made a prediction).
• It adds the flexibility of modern AI (it can handle curves and weird patterns).
• It keeps the speed of the old models (it doesn't require a supercomputer to run).

In short, it's like upgrading a reliable sedan to a sports car that still runs on regular gas. It handles the bumps in the road better, goes faster, and doesn't require a new engine. For economists and financial analysts, this means better predictions, faster decisions, and less time waiting for computers to crunch the numbers.

Technical Summary

1. Problem Statement

Classical ARIMA and SARIMA models remain the standard for time series forecasting in economics and finance due to their interpretability and established statistical theory. However, they suffer from two primary limitations:

1. Rigid Linearity: They assume a strictly linear relationship between current values and past lags (autoregressive terms) and past innovations (moving average terms). This fails to capture nonlinear dynamics common in macroeconomic data (e.g., threshold effects, asymmetric mean reversion) and financial returns.
2. Computational Bottleneck in Rolling Windows: Modern applications (e.g., high-frequency trading, real-time central bank monitoring) require frequent model refitting on sliding windows. Classical ARIMA estimation relies on iterative Maximum Likelihood Estimation (MLE), which involves solving non-linear optimization problems. This process is computationally expensive and becomes a bottleneck when models must be re-estimated thousands of times per second or across thousands of parallel series.

Existing alternatives, such as non-parametric splines or hybrid neural networks, often sacrifice the interpretability of the ARIMA structure or lack closed-form estimators, making them unsuitable for fast, rolling deployment.

2. Methodology: Galerkin–ARIMA and Galerkin–SARIMA

The authors propose Galerkin–ARIMA and Galerkin–SARIMA, a framework that replaces the fixed linear lag operators of classical ARIMA with low-dimensional Galerkin basis expansions while preserving the familiar AR–MA decomposition structure.

Core Concept

Instead of modeling the conditional mean as a linear combination of lags ($y_t = \sum \phi_i y_{t-i} + \dots$), the method approximates the autoregressive and moving average maps as smooth functions projected onto a finite-dimensional basis (e.g., polynomials or splines).

• Nonlinear AR/MA: The autoregressive map $f(x_t)$ and moving average map $g(r_t)$ are approximated as:
  $$f(x_t) \approx \Phi(x_t)^\top \beta, \quad g(r_t) \approx \Psi(r_t)^\top \alpha$$
  where $\Phi$ and $\Psi$ are basis evaluation vectors (e.g., linear, quadratic, or sigmoid terms of lagged values/residuals), and $\beta, \alpha$ are coefficients.

Estimation: Two-Stage Least Squares

A key innovation is the closed-form, two-stage least-squares (2SLS) estimation procedure, which avoids iterative optimization:

1. Stage 1 (Autoregressive Projection): Regress the differenced series $y^{(d,D)}_t$ on the AR basis functions $\Phi(x_t)$ (and seasonal counterparts) to obtain preliminary residuals. This estimates the autoregressive coefficients $\hat{\beta}$.
2. Stage 2 (Moving Average Projection): Regress the Stage 1 residuals on the MA basis functions $\Psi(r_t)$ (and seasonal counterparts) to estimate the moving average coefficients $\hat{\alpha}$.

This approach yields closed-form estimators (solving linear systems $X^\top X \theta = X^\top Y$), which are numerically stable and significantly faster than MLE.

Regularization

To address multicollinearity and overfitting when using rich bases (especially with short rolling windows), the authors introduce a Ridge-Regularized variant. This adds Tikhonov penalties to the least-squares objective, stabilizing the normal equations and dampening forecast spikes without sacrificing the closed-form nature of the solution.

Inference

The paper develops block-bootstrap procedures to handle the "generated-regressor" bias inherent in the two-stage estimation and to account for serial dependence. This allows for the construction of valid prediction intervals and hypothesis testing for exogenous factors (SARIMAX extensions).

3. Key Contributions

1. Theoretical Framework:

   • Establishes consistency and asymptotic normality of the unpenalized estimator under weak dependence ($\beta$-mixing).
   • Derives Jackson-type approximation bounds, showing that the method achieves non-parametric convergence rates governed by the smoothness of the underlying dynamics.
   • Proves that in linear regimes, the estimator behaves as an oracle (matching classical ARIMA), while in nonlinear regimes, it strictly dominates linear SARIMA in asymptotic mean-squared prediction risk.

2. Computational Efficiency:

   • Demonstrates that the two-stage closed-form estimator reduces the computational complexity from $O(I \cdot N)$ (where $I$ is the number of MLE iterations) to $O(N)$ for the design matrix construction plus $O(K^3)$ for the linear solve (where $K$ is the basis size).
   • This enables rolling-window re-estimation that is orders of magnitude faster than MLE-based ARIMA, making it viable for high-frequency and large-scale applications.

3. Model Selection and Inference:

   • Proposes using BIC (Bayesian Information Criterion) for data-driven selection of basis sizes.
   • Provides a rigorous bootstrap framework for constructing prediction intervals that account for both parameter uncertainty and the two-stage estimation structure.

4. Experimental Results

The authors evaluated the framework on synthetic data and real-world macroeconomic/financial series (Quarterly US GDP, Unemployment Rate, Daily S&P 500 returns).

• Synthetic Data:

  • Linear Regimes: Galerkin–SARIMA matched the accuracy of classical ARIMA.
  • Nonlinear Regimes: In processes with nonlinear recursions (e.g., logistic maps) or complex trends, Galerkin–SARIMA significantly reduced forecast errors (RMSE/MAE) compared to linear ARIMA.
  • Speed: The method achieved speedups of several orders of magnitude in rolling forecast scenarios compared to MLE-based ARIMA.

• Real-World Data:

  • S&P 500 Returns: All methods performed similarly (as expected for noisy, weak-signal data), but Galerkin–SARIMA was vastly faster.
  • Real GDP: The Ridge-regularized Galerkin–SARIMA outperformed both unpenalized Galerkin and classical ARIMA in RMSE and MAE. This highlights the importance of regularization when basis expansions are ill-conditioned in finite samples.
  • Unemployment Rate: Galerkin–SARIMA showed competitive accuracy with faster refitting times.
  • Interval Estimation: The block-bootstrap prediction intervals provided well-calibrated coverage, often narrower than the conservative intervals produced by standard ARIMA software.

5. Significance and Implications

• Bridging the Gap: The framework successfully bridges the gap between the interpretability and modularity of classical econometrics and the flexibility of non-parametric machine learning. It retains the "AR–MA" operator structure familiar to practitioners while allowing for nonlinear dynamics.
• Operational Viability: By replacing iterative optimization with closed-form linear algebra, the method solves the latency problem in rolling-window forecasting. This is critical for applications like algorithmic trading, real-time central bank policy analysis, and managing thousands of parallel time series.
• Robustness: The inclusion of ridge regularization and bootstrap inference addresses the practical challenges of overfitting and uncertainty quantification in high-dimensional, dependent data settings.

In conclusion, Galerkin–ARIMA offers a theoretically grounded, computationally efficient, and empirically robust alternative to classical ARIMA for modern, data-rich, and potentially nonlinear time series forecasting tasks.