Forecast all marketing investments with machine learning in one click
Tobias Hink
Data Solutions Engineer
Our latest solution is designed to enable everyone with the ability to forecast performance data from any platform and GA4/UA effortlessly. With just a click of a button, you can gain valuable insights into your business’s future performance.
The Forecasting Solution extracts raw data through multiple APIs and performs data transformations to create a unified set of training data that contains all your marketing costs and revenue from any channel. It harnesses the power of Arima+ alongside other robust methods, all within Google BigQuery, to generate a comprehensive forecast.
Forecast with 95% confidence interval
The Power of ARIMA+
ARIMA+ is like a weather forecast for your business data. Just as meteorologists use patterns from past weather data to predict tomorrow’s weather, ARIMA+ uses your past business data to predict future outcomes.
ARIMA+ forecasting is a time series forecasting method that extends the traditional ARIMA (Autoregressive Integrated Moving Average) model by incorporating additional features, such as seasonality, holidays, and external regressors.
Here’s what ARIMA+ does step-by-step:
- Understanding your data’s rhythm: Every business has a rhythm or frequency. Maybe you get more sales during weekends or particular seasons. ARIMA+ identifies this rhythm.
- Handling inconsistencies: If there are missing dates or duplicate entries in your data, ARIMA+ fixes them.
- Spotting outliers: Just as one exceptionally rainy day doesn’t define a season, a sudden spike in website traffic or sales doesn’t define your general business trend. ARIMA+ identifies and manages these unusual spikes and drops.
- Adjusting for holidays: Holidays can greatly affect business data. ARIMA+ recognizes these patterns and adjusts them.
- Identifying trends: ARIMA+ examines whether your business data is growing, declining, or staying the same over time.
Predicting the future
Our solution automatically creates a comprehensive historical set of training data that always includes the most recent data from GA4 and any advertising platforms. It divides this data into different time series, ensuring accuracy and relevance in your forecasting.
Each time series consists of any defined market- and channel grouping, making it possible to forecast anything from a market as a whole or down to a specific campaign.
The main use case is to forecast short-term performance (up to 90 days), identify deviations from expectations, and take timely corrective action to ensure goals are met.
30 days forecasting solution with ARIMA+
The problem with seasonality
Arima+ solves a well known problem in regression based forecasting with “Holiday Regions”, and even allows for custom holidays not listed in Google’s standard events.
The in-built holidays accounts for shifting holiday dates year-on-year, making it extremely valuable for short peak periods, like Black Friday.
The model is only as good as its training data, but even with just 400 days of training data, it’s excellent at finding patterns in seasons and holidays. In our solution we always provide a minimum of 400 days of training data from the first run, but it’s possible to modify the lookback to finetune the model if needed.
Multiple forecasting methods for data-driven decisions
We understand that one size doesn’t fit all when it comes to forecasting. Our solution offers two complementary forecasting methods using a more traditional and simple approach, enabling users to make data-driven decisions based on comprehensive analysis rather than relying on a single approach. These additional methods include:
Traditional “Run Rate” method: This method provides a straightforward approach but does not account for seasonality. It’s ideal for quick insights into overall trends.
Heuristic method: Using historical data and trend analysis, this model offers an intuitive way to make informed decisions, particularly useful when seasonality is a factor.
Forecasting with multiple methods
Validating model accuracy through backtesting
To ensure accuracy and transparency in our forecasting model, we employ backtesting. This process involves evaluating the model’s predictions using historical data from the previous 38 days, allowing us to compare its forecasts against actual outcomes. Such backtesting provides a statistically sound benchmark for the model’s predictive capabilities. Additionally, incorporating other forecasting methodologies like Run Rate and Heuristic, coupled with current business trend insights, creates a robust framework for a more accurate assessment of the model’s performance.
Testing the model against actual data (Holdout)
Interested in a demo or trial?
Our solution, powered by BigQuery’s ARIMA+ model, is tailored to streamline the forecasting process and provide you with valuable insights to steer your business strategy.
If you are interested in learning more or want to try it out for your business, please fill out the form on our website.
