Forecasting is a common and crucial task that organisations rely on for planning and decision-making. In collaboration with British Airways, this project examines how modern machine learning techniques can be used to improve dynamic forecasting using large-scale business datasets.
Explaining the science
The project investigates the use of Bayesian methods for forecasting and incorporating heterogeneous information.
Bayesian statistics is a principled way of incorporating past experience, domain knowledge of the underlying system, recent trends and new data. Starting with a simple model you iteratively add information and complexity to make incremental improvements in forecasting accuracy. The result is an interpretable model that can be interrogated.
The project falls under the Turing's mission to develop and to promote applications of data science techniques to real-world problems.
The unique challenges addressed by this project are modelling large datasets with complex time series structure and causal effects in a domain where forecast success metrics are commonly hard to define.
The complexity of forecasting problems in the airline industry make it a good case study for testing and developing data science methods. The main challenges in this project range from defining the research question in a complex domain and specifying a success metric to evaluating Bayesian models on large datasets.
The methods and workflows developed in this project are applicable to other large scale data science projects at the Turing.