This is the final part of a three-part blog series. You can read the second article here.

Our previous article explored various types of time series data and the modelling techniques used to analyze them. Building upon that foundation, we'll delve into remaining powerful modelling techniques.

Autoregressive Integrated Moving Average (ARIMA) Time Series Model

ARIMA models combine autoregressive, differencing, and moving average components to capture complex patterns in time series data. ARIMA models, denoted as ARIMA(p, d, q), where:

  • p represents autoregressive lags.
  • d represents the order of differencing.
  • q represents moving average lags.

The autoregressive component (p) means that each future value is based on its previous values, creating a memory effect in the model. For instance, if yesterday's sales were high, there's a good chance today's sales will also be high. The integrated part (d) refers to differencing. Differencing involves subtracting each value from its previous value to make the series more stationary. Stationary data is easier to work with and analyze. It means the data has no long-term trends or patterns that change over time. The moving average component (q) smooths out random fluctuations in the data to identify underlying trends. It helps capture short-term changes and patterns in the data.

ARIMA models are versatile and widely used for forecasting and modelling non-stationary data. Let's say you're analyzing monthly sales data. ARIMA would use past sales figures (autoregressive), apply differencing to remove trends or seasonality and smooth out random fluctuations with a moving average. This helps predict future sales accurately.

Auto-correlation Functions (ACF) and Partial Auto-Correlation Functions (PACF) are essential tools in time series analysis, particularly when dealing with autoregressive models like ARIMA.

The Auto-Correlation Function (ACF) is like a magnifying glass for time series data. It helps us see how much each data point is related to its past versions at different time intervals. When you calculate the ACF, you get a bunch of numbers, each telling you how strong the relationship is at different points in time. These numbers can range from -1 to 1. A value of 1 means a perfect match between today's temperature and yesterday's. -1 means they're perfectly opposite, and 0 means there's no connection at all. ACF helps us spot patterns or trends in the data.

You can think of the Partial Auto-Correlation Function (PACF) as a sharper tool than the ACF. While the ACF gives us a general idea of how today's data point relates to its past, the PACF zooms in on the direct relationship, ignoring any middlemen. Imagine you're tracking daily sales data. The PACF would tell you exactly how yesterday's sales directly impact today's without considering any other days in between. This makes it super helpful for figuring out how many past observations we need to predict future values accurately, especially in models like ARIMA.

So, in simpler terms, while ACF gives a broader view of relationships over time, PACF provides a focused view, helping us pinpoint direct influences between data points. These tools provide insights into data patterns, guide model parameter selection, aid in preprocessing steps, and contribute to improved forecasting performance. Therefore, understanding ACF and PACF lays a solid foundation for implementing ARIMA models and extracting meaningful insights from time series data.

Seasonal ARIMA (SARIMA) Time Series Model

SARIMA is like a big brother to ARIMA. It takes everything ARIMA does and adds a special twist to handle seasonal patterns in the data. Just like ARIMA, SARIMA looks at past data to predict future values. But it also considers seasonal factors, like spikes in sales during the holidays. ARIMA handles both short-term seasonal changes and long-term trends in the data, providing a complete picture of the patterns. SARIMA models provide a comprehensive framework for capturing complex temporal patterns in the data.

Vector Autoregression (VAR) Model

VAR models extend the concept of autoregression to multiple time series variables.

Each variable in a VAR model is regressed on its own past values (lags). This means we look at how each variable depends on its own past behaviour. For example, if we're looking at GDP, we might regress it on its GDP values from previous periods to see how it changes over time.

In addition to regressing on its own past values, each variable is also regressed on lagged values of other variables in the system. This means we're considering how changes in one variable might influence the others. For example, if we're looking at GDP, we might regress it on past unemployment or inflation rate values to see how these variables affect GDP.

By considering the past behaviour of each variable and the interactions between variables, VAR models help us understand how different factors affect each other over time. For instance, in economics, we can use VAR models to see how changes in one economic indicator, like interest rates, affect other indicators, such as investment, inflation, or exchange rates.

VAR models capture the dynamic interactions between variables over time. This means that changes in one variable can affect the behaviour of other variables in the system, leading to a feedback mechanism. For example, in macroeconomic forecasting, changes in GDP might influence consumer spending, which in turn affects inflation rates and unemployment. A VAR model can capture these interdependencies and their lagged effects on each other.

One key advantage of VAR models is their ability to forecast multiple time series variables simultaneously. Instead of creating separate models for each variable, VAR models estimate the relationships between all variables in the system simultaneously. This simultaneous estimation allows for a more comprehensive understanding of the interactions between variables and can lead to more accurate forecasts than univariate models.

These modelling techniques and appropriate diagnostic tools, such as ACF and PACF, are crucial in analyzing time series data and identifying patterns.


Understanding the characteristics and implications of each type of time series data is crucial for selecting appropriate analytical techniques and building accurate forecasting models. Whether identifying trends, capturing seasonal variations, or modelling complex cyclical patterns, mastering the nuances of time series data types empowers analysts to extract meaningful insights and make informed decisions.