REPOSITORY > RESULTS

Abstract

This thesis examines how machine learning can be applied in demand forecasting. In particular,
it describes a novel approach toward lumpy and intermittent demand forecasting.
It advocates using a two-fold model for forecasting lumpy (irregular demand occurrence,
strong demand size variability) and intermittent (irregular demand occurrence, little demand
size variability) demands when considering fixed forecasting horizons. The two-fold
model comprises a classification model to predict demand occurrence and a regression
model to predict the demand quantity. By dividing the forecasting problem into these two
dimensions, insights into the reasons affecting the forecast quality are exposed by allowing
us to understand to what extent the forecasting model fails to predict demand occurrence
or demand quantity. Furthermore, a new set of metrics is proposed to assess the quality
of the forecasts: ROC AUC to determine the quality of the classifier, variations of the
MASE metric to measure the quality of the regression component, and the SPEC metric
to determine the impact of the forecast on the stock costs.
While an accurate prediction of when the demand event occurs is of utmost importance,
a prediction score alone does not provide an intuitive interpretation to the human.
Furthermore, multiple models produce different predictive score distributions. Therefore,
an additional model (a.k.a. calibrator) can be trained to map the original predictive scores
to probabilities (in our case, the probability of demand occurrence). While much research
has been devoted to measuring the quality of such calibrators, it is assumed that ground
truth is required to measure and calibrate machine learning models. This thesis presents a
novel approach to measuring the quality of the calibration, addressing the shortcomings of
a widely adopted metric. The results show that the newly proposed metrics can estimate
the calibrators’ quality without using ground truth when no concept drift occurs. This
approach could be helpful in at least two scenarios: (i) monitoring the performance of the
calibrator over time without incurring labeling costs and (ii) training a calibrator without
ground truth (human label on whether the event under consideration took place or not).
While some experiments were performed on training a calibrator without ground truth,
the first results were not promising, and more research must be invested in this direction.
Finally, the thesis highlights a use case of demand forecasting where explainable artificial
intelligence was used to determine features that were relevant to a given forecast.
Furthermore, such insights were enriched with expert knowledge from ontologies and insights
obtained through text-mining techniques to build high-level explanations leveraging
information retrieved from external sources.