October 20, 2024
Choosing The Best Time Series Forecasting Model For Your Use Case
When trying to predict the future based on past data, choosing the right forecasting model can feel scary, especially if you're not a data scientist. But don’t worry, you don’t need to be an expert to pick a forecasting model that works for you.
5 min read
Forecasting
Whether you’re predicting sales, website traffic, or the weather, each forecasting model has its strengths. This guide will break down the models by how much computing power they need (how fast they run) and when they are the best fit for your data. We’ll use real-world examples to make this easy to understand.
How to Choose the Right Forecasting Model
When deciding on a forecasting model, here are a few questions to ask:
- Does your data show a simple trend over time (going up or down)?
- Models like Linear Regression and ARIMA are great for simple trends where data gradually increases or decreases.
- Do you see repeating patterns, like sales going up every December?
- Holt-Winters, SARIMA, or even Prophet are better when your data has clear seasonal patterns.
- Are there outside factors that affect your data?
- If external factors (like weather, marketing spend, or holidays) affect your results, models like SARIMAX or machine learning models (e.g., XGBoost, Random Forest) can handle these.
- Do you have a large dataset with complex patterns?
- For more complicated data that doesn’t follow simple trends or patterns, newer models like NeuralProphet or Nixtla's NeuralForecast use advanced techniques like neural networks to predict.
- Do you need to forecast for different groups, like regions or stores?
- Hierarchical forecasting is best when your data is grouped (e.g., store-level and region-level sales).
Table of Forecasting Models by Compute Resources and Use Case
Here’s a table to help you understand which model might be right for you based on how complex your data is and how much time/resources it will take to run the forecast.
| Model | Model Type | Additional Data Needed | Compute Resources | Best Used When | Accuracy |
|---|---|---|---|---|---|
| Linear Regression | Basic Statistical | ds, y, optional extra factors (regressors). No holidays. | Very Low | Your data has a clear straight-line trend over time. | Low |
| ARIMA | Statistical | ds, y. No extra factors or holidays. | Low | The data doesn’t have any seasonal patterns and changes slowly. | Medium |
| Holt-Winters (Exponential Smoothing) | Statistical | ds, y. No extra factors or holidays. | Low-Medium | The data shows clear repeating patterns (seasons) over time. | Medium |
| SARIMA | Statistical | ds, y, seasonal information (e.g., length of seasons). No holidays. | Low-Medium | Data has both trend and seasonal patterns (e.g., yearly sales patterns). | Medium-High |
| SARIMAX | Statistical | ds, y, extra factors (e.g., marketing spend). No holidays. | Low-Medium | The data is affected by outside factors like weather, marketing, or other variables. | Medium-High |
| StatsForecast (Nixtla) | Statistical | ds, y, optional extra factors. Holidays supported. | Low-Medium | You need a fast, accurate forecast for lots of data and want to use simple statistical methods. | High |
| MLForecast (Nixtla) | Machine Learning | ds, y, optional extra factors. Holidays treated as extra factors. | Medium | You have a lot of data and want to use a more powerful machine learning model to predict the future. | High |
| AutoML (Google, Azure) | Machine Learning | ds, y, optional extra factors. Holidays treated as extra factors. | Medium | You want an automated tool to pick the best model for you without needing to adjust anything manually. | Varies |
| Random Forest | Machine Learning | ds, y, optional extra factors. Holidays treated as extra factors. | Medium | Your data has complex patterns, and the future is influenced by many factors. | High |
| XGBoost | Machine Learning | ds, y, optional extra factors. Holidays treated as extra factors. | Medium | You need fast, accurate predictions, and the future is shaped by many factors. | High |
| LightGBM | Machine Learning | ds, y, optional extra factors. Holidays treated as extra factors. | Medium | You have a large amount of data and need faster predictions. | High |
| Amazon Forecast (Chronos) | Machine Learning | ds, y, optional extra factors and holidays. | Medium-High | You want to make predictions on a big dataset and don’t mind using a service like Amazon to do it. | High |
| Prophet | Bayesian Forecasting | ds, y, optional extra factors, holidays. | Medium | You want an easy-to-use model that handles repeating patterns and holidays well. | Medium-High |
| NeuralProphet | Neural Network | ds, y, optional extra factors and holidays. | Medium-High | Your data shows repeating patterns and spikes (e.g., sales driven by events) and needs more flexibility. | High |
| NeuralForecast (Nixtla) | Neural Network | ds, y, optional extra factors. Holidays treated as extra factors. | High | Your data is complicated with lots of ups and downs, and you want to use advanced models. | High |
| HierarchicalForecast (Nixtla) | Hierarchical | ds, y, grouped data (e.g., by region, store). Holidays treated as extra factors. | High | You have groups of data (e.g., sales by store/region) and need to predict for each group. | High |
| LSTM | Neural Network | ds, y, optional extra factors. Data must be prepared in sequences. | High | You have long sequences of data with lots of changes over time and want to predict far into the future. | High |
Example Use Cases
- Simple Trend: Let’s say you run a coffee shop and want to predict how many cups of coffee you’ll sell next week. If your sales are steadily increasing week after week, then Linear Regression or ARIMA might be enough to forecast that simple trend.
- Repeating Patterns: If you own a toy store and sales skyrocket every December, you’ll want a model that handles seasonality well, like Holt-Winters or Prophet, which are great for data with clear seasonal patterns.
- External Factors: If you run an online business, your website traffic might be affected by things like ad spending, social media campaigns, or even the weather. Models like SARIMAX, Random Forest, or XGBoost can include these external factors in the forecast.
- Multiple Locations or Groups: If you manage sales across multiple stores, you might want a model like HierarchicalForecast, which lets you forecast sales at both the store and regional levels.
Choosing the right forecasting software and model depends on your data’s complexity, the patterns you see, and the external factors that affect your results. Simpler models work well when your data trends in a straightforward way, but when things get complicated with seasonality, external events, or group data, more advanced models are the way to go.
With this guide, you should be able to select a model that fits your needs without getting bogged down in technical jargon. Happy forecasting!