ebook
The Battle for ATTENTION: Will AI Save The Advertising Industry in a Privacy-Driven Economy?

3 Powerful Time-Series Analysis Techniques to Drive Better Insights

By SCUBA Insights

Time-series data is everywhere—whether or not your brand is equipped to handle it. Data-driven organizations need time-series analysis platforms to make the most of their data, but some brands may not realize there are different techniques for achieving time-series analysis. The question isn’t whether time-series analytics platforms are worth it—they are—but knowing which analysis technique is best suited for your brand goals and needs.

 

In today’s blog, we’ll explore three common time-series analysis techniques and how they can be leveraged to solve common business problems across industries.

 

3 common methods to perform time-series analysis

In time-series analysis, data points are recorded at consistent intervals over a set time period. This allows brands to compare data points against different variables and times, making it easier to discern trends and outliers.

 

However, there is no universal method to perform a time-series analysis, and some methods may be better suited for certain data sets than others.

 

  1. 1. Univariate Box-Jenkins Models

Univariate Box-Jenkins modeling, or autoregressive integrated moving average (ARIMA), is used to forecast a single time-dependent variable. Box-Jenkins models come in two forms: univariate and multivariate. As their names suggest, univariate models forecast a single variable, whereas multivariate models forecast multiple variables.

 

We will explore multivariate models in the next section, so let’s quickly break down the mechanics of univariate modeling. 

 

First, the following values must be determined:

 

  • (p) = the  number of autoregressive terms

 

  • (d) = the number of non-seasonal differences required to make data stationery. This means the properties of the time series are consistent, regardless of when the series was observed.

 

  • (q) = the number of lagged forecast errors in a prediction equation.

 

Once determined, these values are plugged into the following models:

 

  • General autoregressive model - AR(p): predicts using previous values of the time-dependent variable.

 

  • Integrated I - (d): The difference is taken d times until the original series becomes stationary. A stationary time series is one whose properties do not depend on the time at which the series is observed.

 

  • General moving average model - MA(q): predicts using the series mean and previous errors

 

By combining previous time-series data, as well as means, errors, and seasonality, ARIMA can predict a single, time-dependent variable.

Benefits of univariate ARIMA models

  • Accurate short-term forecasting: ARIMA excels at short term-forecasting. They produce the most accurate forecasts within a 12 to 18-month period.

 

  • Operates exclusively off historical data: Since ARIMA models operate exclusively off historical data, they require less data scrubbing to become functional, allowing you to model faster.

 

  • Generalizes non-stationary data: Though using stationary data simplifies the ARIMA forecasting process, determining (d) is relatively simple. The result is stationary data that can easily be applied to other more complex models.

Challenges with univariate ARIMA Models

  • Poor long-term forecasting: Because general autoregressive modeling operates off previous values, forecasting accuracy can drop off past eighteen months.

 

  • Difficulty adapting to sudden changes: ARIMA models are linear, meaning that they do not account for sudden changes, even with seasonality accounted for.

 

  • Vulnerable to poor calculations: ARIMA is only as good as its data. If historical data is input incorrectly or differencing values are miscalculated, ARIMA models are practically useless.

When to use univariate ARIMA models

Univariate ARIMA models are best used to understand a single, time-dependent variable, such as temperature over time or investment returns over time.

 

In 2020, the National Library of Medicine published a case study on the use of Box-Jenkins models to forecast COVID-19 spikes in heavily afflicted countries. Using time-series data provided by the World Health Organization, epidemiologists accurately predicted that while Spain and the United States would experience COVID-19 spikes, cases would drop in China. This information helped these and other countries adapt their COVID-19 strategies—all thanks to time-series analysis.

2. Multivariate Box-Jenkins models

Multivariate Box-Jenkins models, or Autoregressive Moving Average Vector (AMAV) models, function similarly to their univariate counterparts with one key difference: they can be used to model more than one variable at once.

 

Many different techniques fall under the multivariate box Jenkins umbrella, but all generally split into two camps: dependent and independent techniques.

 

  • Dependent techniques: Dependent modeling techniques examine causality between data points. If the values of two or more data points can be used to explain, describe, or predict the value of another, dependent variable, they are considered dependent.

 

  • Multiple regression: An extension of simple linear regression used to predict the value of a variable based on two or more variables.

 

  • Conjoint analysis: A survey-based technique that correlates different user values (features, functions, benefits, etc.) and end-user decisions.

 

  • Multiple discriminant analysis: Determines groups within data sets by finding linear combination variables.

 

  • Linear probability model: A regression model that predicts binary outcome variables (0 or 1, yes or no) using one or more explanatory variables.

 

  • Canonical correlation analysis: Summarizes linear relations between two sets of variables.

 

  • Structural equation analysis: A broad, versatile framework that analyzes structural techniques. 

 

  • Independent techniques: If no variables are dependent on each other, independence modeling techniques can be used to understand the structural makeup and underlying patterns within the dataset.

 

  • Factor analysis: a pre-processing technique that prepares data for other models by condensing multiple variables together.

 

  • Cluster analysis: Used to classify data points into relative groups, or “clusters.”

 

  • Multidimensional Scaling: Displays the relative distance of data points using a table of distances. This table, called a proximity matrix, may consist of multiple dimensions.

 

  • Correspondence Analysis: Visualizes rows and columns of data as points on a map.

Benefits of multivariate AMAV models

Multivariate AMAV techniques share the same benefits as ARIMA, in addition to the following:

 

  • Tracks relationships between data points and variables: The relationship between data and variables is hardly linear. Multivariate AMAV techniques take this into account, allowing you to adjust variables to better understand the relationship between them.

 

  • Greater accuracy than univariate models: Assuming your data is correct, multivariate models paint a more comprehensive picture of your data. By considering all the variables that could affect your data, you are less likely to miss something or make an incorrect assumption.

Challenges with multivariate AMAV models

Multivariate AMAV models face the same challenges as univariate AMIRA models, but modeling additional variables can present additional challenges:

 

  • Larger data requirements: As more variables are introduced, AMAV techniques require more data to derive accurate conclusions. All of this data must be cleaned so it can be input into an AMAV model, which can cause a serious time and resource drain.

 

  • Higher computing power requirements: The computing prowess required for Multivariate AMAV is more complex than its univariate ARIMA counterpart, so this technique may not be ideal for brands already struggling with strained bandwidth.

When to use multivariate AMAV techniques

Multivariate AMAV techniques are best used with large data sets to find correlations between multiple variables, such as discovering dependence between variables or predicting future relationships. These techniques can also be used to transform data for other models by simplifying data structures or grouping variables together.

 

Dependent multivariate techniques, such as conjoint analysis, are often utilized by marketing teams to make sense of survey-based qualitative data. By correlating data points to end-user decisions, marketers can learn more about their users and predict the success of future campaigns.

 

Banks and other financial institutions can leverage independent multivariate techniques, such as cluster analysis, to predict future behavior. Suspicious credit card purchases often fall outside predictable user behavior, allowing banks to flag accounts quickly and protect their customers.

3. Holt-Winters exponential smoothing

The Holt-Winters exponential smoothing technique is a simple but effective alternative to the Box-Jenkins. Instead of computing entire data sets, Holt-Winters uses exponential smoothing to predict typical present and future values.

 

Holt-Winters is sometimes referred to as triple exponential smoothing as it combines the following three smoothing methods:

 

  • Exponential Weighted Moving Average (EWMA) method: EWMA is the foundation of Holt-Winters, as the exponential weighted moving average is what “smooths” a time-series data. This is achieved by finding the central value between data points.

 

  • Holt’s Smoothing Method: By augmenting a simple exponential smoothing model with a 2nd EWMA to account for slope, Holt’s smoothing method allows for linear forecasting.

 

  • Winters Smoothing Method: The addition of a third EWMA factors seasonality into the forecasts.

 

To begin modeling, the following parameters must be established.

 

  • (a) - smoothing parameters, or the “weight” of each exponentially decreasing data point

 

  • (b) - length of a season

 

  • (y) - number of periods in a season

 

Once these parameters are applied, Holt-Winters can predict complicated seasonal trends by discovering the central value and factoring in slope and seasonality.

Benefits of Holt-Winters Exponential Smoothing

  • Less data-intensive: Holt-Winters is a simple model that does less with more. Compared to box-Jenkins models, Holt-Winters requires far fewer data to achieve the same level of accuracy.

 

  • Reduced computing power: A softer data requirement means less computing power is required for accurate forecasts.

 

  • Easily identifiable outliers: Outlier data points are immediately identifiable against a Holt-Winters generalized trend forecast.

Challenges of Holt-Winters Exponential Smoothing

  • Poor performance within shorter time frames: Multiplicative seasonality can be a serious issue within small seasonal time frames, which may result in inaccurate generalized values.

 

  • Higher skill requirement: Since Holt-Winters requires less data or computing power, great care must be taken when building the model. Accuracy can vary significantly with even the smallest parameter changes, so this technique may not be a good fit for less experienced teams.

When to use Holt-Winters Exponential Smoothing

The versatility and relative simplicity of Holt-Winters exponential smoothing makes this technique an effective choice for many brands. Smaller brands may appreciate the less intensive data and computing requirements, but larger brands can also leverage Holt-Winters to forecast generalized trends and identify outliers.

 

Accurate forecasting is critical for practically every industry, but retail and distribution brands can make especially good use of Holt-Winters by forecasting typical inventory needs and adjusting order sizes based on seasonality.

Problem-solving with time-series analytics techniques

1. How can telecom brands be more proactive about 5G data security?

When it comes to telecom data security, 5G networks are a catch-22. While 5G offers enhanced security features 2G, 3G, or even 4G couldn’t handle, its decentralized networks and rapid proliferation have resulted in an explosion of vulnerable access points. With so many variables to track, from third-party vendors to AI vulnerabilities, it’s no wonder Deloitte's 2022 industry report found 80% of telecom executives say 5G data security is a top concern.

 

Solution: Independent multivariate Box-Jenkins technique

 

Fortunately for telecom brands, this time-series technique is tailor-made to track multiple variables such as third-party vendors or access points.

2. How can brands gauge SaaS purchase ROI?

SaaS usage exploded in 2022. A recent SaaS industry report found SaaS usage increased by 57% in the last year among US and European organizations. Unfortunately, that growth may not be sustainable. A recent SaaS industry report revealed a few startling statistics. Nearly a third of brands believe they waste up to 40% of their SaaS budget on underutilized subscriptions; 49% of brand leaders say controlling application sprawl is a major challenge.

 

Solution: Univariable Box-Jenkins technique

 

A univariate Box-Jenkins model is more than sufficient to determine which SaaS brands provide the best bang for your buck. Not only will the model present historical usage data, but growing brands can leverage Box-Jenkins to forecast which SaaS purchases will provide higher ROI as your teams’ adoption increases.

3. How can media brands optimize subscriber monetization?

A new phase of the streaming wars has begun. As subscriber numbers plateau, monetizing existing subscribers is just as important as securing new ones. However, most subscribers have become accustomed to ad-free viewing, so brands must tread lightly or risk driving their customers to the competition. Now more than ever, media brands must strike a fine balance between paid tiers, increased ad cadence, and perceived value.

 

Solution: Dependent multivariate Box-Jenkins technique

 

Dependent multivariate Box-Jenkins models are ideal for finding the correlations between qualitative data, such as customer satisfaction, and quantitative data, such as time-on-site. Rather than blindly increasing ad cadence and hoping for the best, this technique can help media brands forecast subscriber behavior before implementing new ad tiers.

4. How can mobile gaming brands quickly identify in-app purchase vulnerabilities?

Mobile gaming revenue is expected to reach $152 billion in 2022, a staggering 72% of which comes exclusively from in-app purchases. With so much at stake, quickly identifying exploits such as money or item duplication is a top priority. Still, mobile gaming brands must tread lightly, as blunt force approaches such as insta-banning suspicious accounts could drive away legitimate users.

 

Solution: Holts-Winters exponential smoothing technique

 

Mobile gaming brands can leverage Holt-Winters models to forecast typical user inventories based on time-in-app and money spent on in-app purchases. Suspicious accounts–such as users with locked items or impossibly massive inventories– would stick out like a sore thumb. Once these accounts are identified, brands can learn how they exploited the game, patch vulnerabilities, and stymie revenue loss.

Perform time-series analysis with Scuba

Time-series forecasting might be complicated, but choosing an analytics solution doesn’t have to be. Scuba’s real-time customer intelligence platform empowers brands with the insights they need to adapt to today’s rapidly evolving landscape–regardless of data set size, variables, or season length.

 

  • Eliminate lengthy ETL processes with fully scalable, raw data ingestion
  • Glean insights without straining your IT teams with our no-code queries
  • Democratize your data with our intuitive, non-technical UI
  • Reduce time-to-insight from weeks to minutes

 

Want to learn more about time-series analytics techniques? Request a demo today or talk to a Scuba expert.

Stay Updated

Stay in touch with Scuba with fresh insights delivered to your inbox.

Ready to dive in?

We'd love to connect.

Talk to an Expert