Balancing Technology and Business Acumen in Transaction Volume Forecasting

I have been experimenting with various forecasting and inference Python libraries the past few weeks, and I have reinforced some of my thoughts regarding forecasting transaction volumes using these libraries.

No single time series forecasting tool can perfectly predict transaction volumes for your contact center. Out-of-the-box libraries are straightforward to use and can adequately forecast a business's transaction patterns. However, they require users to adjust hyperparameters such as changepoints, holiday effects, seasonality and other parameters.

Even after fine-tuning these hyperparameters, the model will not forecast transaction volumes flawlessly. This is because it cannot anticipate future business changes that were not part of the training data. Businesses evolve, introducing new products and services, merging with or acquiring other companies, and making various changes that affect future transaction volumes. Thus, while the models can establish a baseline, accurate forecasting still necessitates human input and judgement to refine these predictions.

There are numerous inference libraries that evaluate the relationships between independent variables fed into the model and output a dependent variable, like transaction volume. These models are valuable for understanding the influence of various factors. However, its bit of a catch 22  arises when using inference libraries to predict future transaction volumes. Accurate predictions require forecasted independent variables, making downstream forecasts heavily reliant on upstream predictions. This creates a scenario where inference forecasts are essentially based on other forecasts, increasing the potential variance in outcomes.

To be a proficient forecaster, one must be technically skilled in using these models and have a thorough understanding of the business to make informed adjustments. Rather than relying on a single model, it is advantageous to use multiple models in competition, selecting several to provide a range of forecasts. This approach allows for the creation of various scenarios to effectively plan and test the capacity of contact centers under different conditions.