Why Include Lags? The Key Reasons in Autoregressive Models

In the world of economics, predicting the future isn’t about looking into a crystal ball; it’s about looking closely at the past. Economic and financial data-things like GDP growth, inflation rates, or stock volatility-don’t behave randomly. They exhibit a kind of historical echo. This phenomenon demands a specialized modeling approach, and few are as foundational as the Autoregressive (AR) model. At its heart, the AR model is a sophisticated time-traveler, using a variable’s past values to forecast its present and future state. But here lies the critical question: how far back should we look? This is where the concept of ‘lags’ comes into play. Determining the correct number of lags ($p$) in the AR model, represented in the equation $Y_t = \mu + \phi_1 Y_{t-1} + \phi_2 Y_{t-2} + \dots + \phi_p Y_{t-p} + \xi_t$, is not just a statistical formality; it is the key to creating a model that truly understands the dynamics of the economy.

The core purpose: diagnosing and capturing persistence 🔗

The primary reason for including lagged values of the dependent variable ($Y_{t-1}, Y_{t-2}$, etc.) in an AR model is to mathematically capture persistence. Persistence, sometimes called ‘serial dependence’ or ‘autocorrelation,’ describes the tendency of a time series variable to stick close to its previous values. Think of an inflationary shock: if prices jump today, they won’t typically fall back to baseline tomorrow. The high price level persists, influencing the price level next month, and the month after. This “stickiness” is the defining feature of most real-world economic data.

Understanding economic “stickiness” 🔗

When an economic variable is persistent, it means that a shock (like an unexpected trade war or a major policy change) has effects that decay slowly over time. If we were to use a simple linear regression that only included external factors ($X_t$), we would be implicitly assuming that any past effects vanish immediately, which is rarely the case in macroeconomic phenomena. For instance, consider the market capitalization of a company following unexpected bad news. The immediate dip is only part of the story. The full effect may be “distributed” over several months as the company adjusts its production and inventory, a concept often explored in distributed lag models that informs our understanding of AR structure [1]. By incorporating the variable’s own lag, $Y_{t-1}$, we ensure that the model acknowledges that today’s value ($Y_t$) inherits some momentum from yesterday’s value.

The crucial role of autocorrelation 🔗

The choice of how many lags, $p$, to include is directly dictated by the data’s autocorrelation structure. Autocorrelation measures the correlation between a time series and a lagged version of itself. An AR model is essentially trying to explain the current value using the autocorrelation information embedded in the past values. If we find that $Y_t$ is highly correlated with $Y_{t-1}$ (Lag 1) and moderately correlated with $Y_{t-2}$ (Lag 2), but negligibly correlated with $Y_{t-3}$ or $Y_{t-4}$, then an AR(2) model is likely sufficient. Ignoring a statistically significant lag will lead to a misspecified model, meaning our model residuals ($\xi_t$) will contain the dependency we failed to model. This results in biased and inconsistent parameter estimates, violating the assumptions of Ordinary Least Squares (OLS) estimation when applied to dynamic models [2]. Econometricians rely on tools like the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) plots to visually diagnose exactly how many lags are necessary to adequately mop up this serial dependence.

Incorporating memory effects: the power of p 🔗

The order of the autoregressive process, $p$, is often analogized to the model’s “memory.” A model with a higher $p$ has a longer, more detailed memory of its own past. This memory is crucial because the true underlying process generating the economic data (the “data-generating process” or DGP) rarely follows a simple AR(1) structure.

The difference between short and long memory 🔗

Most standard AR models (and their generalizations, ARMA models) are considered short memory processes. This means that the influence of a shock, though persistent, decays relatively quickly-typically at an exponential rate. In the continuous-time analogy, an AR(1) process with a coefficient ($\phi_1$) less than one describes a system that exponentially decays towards an equilibrium point [3].

However, the number of lags we include directly shapes the length and complexity of the model’s memory. When we move from an $AR(1)$ to an $AR(p)$, we are allowing for complex dynamics, such as cyclical or oscillatory behavior. For example, a second-order AR(2) model can generate wave-like fluctuations resembling business cycles, whereas an AR(1) can only produce smooth decay. A higher-order AR model can be re-expressed as a set of first-order differential equations, demonstrating its capacity to capture rich, multi-period dynamics.

When one lag just isn’t enough (the p-order system) 🔗

In the classic AR(1) model ($Y_t = \mu + \phi_1 Y_{t-1} + \xi_t$), the relationship is direct: today is only affected by yesterday. But what if there’s a three-month lag in monetary policy transmission, or a four-quarter lag in business investment decisions? In India, for example, the impact of the monsoon on rural demand might not fully manifest in corporate earnings until two or three fiscal quarters later.

If the true DGP is an AR(3), and we incorrectly specify an AR(1), the coefficients we estimate will be severely flawed because the model attributes the lingering influence of $Y_{t-3}$ to the current error term, $\xi_t$. By including multiple lags, we allow the model to capture the entire historical dependency chain that is inherent in the data. An AR(p) model says: “The variable $Y$ at time $t$ is a linear combination of its values over the last $p$ periods, plus a completely random, unexplained shock.” This rich, structured memory is essential for accurate forecasting, as a one-time shock will propagate through the system for $p$ periods, or even infinitely into the future if the process is stationary (albeit decaying) [4].

Accounting for time dependency and model validation 🔗

Beyond merely capturing persistence, the inclusion of lags serves a vital diagnostic and structural purpose: ensuring the model is correctly specified so that statistical inference can be trusted. In econometrics, we seek a parsimonious model-one that uses the fewest possible parameters while still explaining the data adequately. However, using too few lags (underfitting) or too many lags (overfitting) each presents its own set of statistical pitfalls.

From simple dependence to dynamic specification 🔗

The coefficients $\phi_1, \phi_2, \dots, \phi_p$ are not mere nuisance parameters; they define the precise nature of the time dependency. In an AR model, these coefficients are allowed to adapt to the historical data, giving us a quantitative measure of how much weight the system places on each past observation. For instance, if $\phi_1$ is large (close to 1) and all subsequent $\phi$’s are small, the variable exhibits very short-term persistence. If $\phi_1$ is small, but $\phi_4$ and $\phi_8$ are large (and statistically significant), the variable likely has a strong seasonal or cyclical component (like quarterly retail sales or annual crop yields).

In advanced applications, sometimes economists employ time-varying autoregressive (TVAR) models, where the coefficients themselves (the $\phi$’s) are allowed to change over time, acknowledging that the underlying economic structure-the way people or businesses react to past information-may itself evolve [5]. For foundational AR models, including enough fixed lags is the minimum requirement to capture any fixed, yet complex, dynamic structure.

Avoiding bias: the diagnostic necessity of lags 🔗

One of the most critical reasons for correctly choosing $p$ is to ensure that the final error term, $\xi_t$, is white noise. White noise means the error term is unpredictable and not correlated with its own past values ($\xi_t$ is independent of $\xi_{t-1}, \xi_{t-2}$, etc.). If the lags in the model do not fully account for the historical dependence, the residual error $\xi_t$ will be autocorrelated (also known as serial correlation).

In dynamic models like the AR model, the presence of autocorrelated errors is devastating because it causes the Ordinary Least Squares (OLS) estimators for the coefficients ($\phi$’s) to be biased and inconsistent. Simply put, the estimation is fundamentally broken, and the results cannot be trusted. Therefore, the common procedure for determining $p$ is to select a model order high enough to make the residuals white noise, often using criteria like the Akaike Information Criterion (AIC) or Schwarz’s Bayesian Information Criterion (BIC), which balance explanatory power against model complexity (penalizing the inclusion of too many unnecessary lags) [6]. We include lags primarily because they are the only mechanism in the AR framework capable of correctly specifying the dynamic feedback structure of economic time series.

What do you think? If a country’s GDP growth (a highly persistent variable) were accurately modeled by an AR(1) process, what specific policy implication might that suggest about the speed of recovery following a severe economic shock? Conversely, what characteristics of economic data might suggest the need for a non-integer, or “fractional,” differencing parameter, indicating “long memory”?

References

1. https://www.mathworks.com/help/econ/time-series-regression-viii-lagged-variables-and-estimator-bias.html
2. https://www.soa.org/49eabf/globalassets/assets/library/newsletters/predictive-analytics-and-futurism/2017/june/2017-predictive-analytics-iss15-lai-lu.pdf
3. https://www.investopedia.com/terms/a/autoregressive-integrated-moving-average-arima.asp
4. https://www.mdpi.com/2571-905X/5/3/46
5. https://www.lsu.edu/business/economics/files/workingpapers/pap97_27.pdf