COVID-19 is causing a widespread change of behaviour. Leaving aside personal beliefs on whether this reaction is justified, it can be seen that this response greatly impacts economic markets and the conduct of everyday business. In more technical terms, one speaks of structural changes, and their effects are non-trivial to those who are forecasting future behaviour. How can impacts like this one – thus, structural changes – be handled in time series analysis?
A Structural Change
Wikipedia defines a structural change as “a shift or change in the basic ways a market or economy functions or operates” (Wikipedia contributors, 2019a). A structural break, however, is defined as an unexpected change over time in the parameters of regression models, which can lead to huge forecasting errors and unreliability of the model in general (Wikipedia contributors, 2019b). Both are often closely related: structural breaks in data are often caused by structural changes in the underlying (economic) processes where the data originate from. In business, unreliable models and large forecasting errors are highly detrimental, since managers tend to base important decisions on such models and their outputs. Therefore, it is vital to timely and successfully combat these issues.
This concept is easy to get behind, but actually executing on it is a more complicated challenge. Structural changes render existing models useless. The obvious response is to build new models, based on new data, to account for these structural changes. This is often a tedious process, as new significant features need to be identified and new models need to be trained. Moreover, doing so by hand takes time, and time might not be available in these situations. AutoML improves this situation by automating the model building process, but still tends to require a considerable amount of time. Additionally, AutoML has no impact on the identification of significant features. However, following such an increase in the required frequency of model building, both the automation of and the time needed for this process are of the upmost importance, especially when one wants to minimise business disruption.
Tangent Works’ model building engine, TIM (Tangent Information Modeller), claims to generate and apply high-quality forecasting models for time series data in just a few seconds, which means TIM overcomes the challenges posed by structural changes (Tangent Works, 2018). How does TIM accomplish this?
Multiple innovative aspects come together to reach the solution. Firstly, in contrast with handcrafted modelling techniques and AutoML, TIM automates feature engineering. This automatic feature engineering includes going through all input variables and understanding which subselection will contribute to the final result. It also includes generating many new artificially created features from the original variables (expansion) which can be useful for model building, and reducing this newly created set of features to a smaller, useful subset (reduction) to achieve model stability and prevent overfitting.
TIM goes even further, unifying this automated feature engineering, model building and model deployment (i.e. application of the model in production) into one single step. The technology that accomplishes this is called RTInstantML (real-time instant Machine Learning). This extensive automation enables users to train a new model for each desired forecast, eliminating the need of designing and building models beforehand while ensuring optimal models are used during times of variable feature availability. RTInstantML thus helps to overcome the challenges of structural changes in forecasting through both ad hoc feature engineering and on-demand model building.
About the author
Elke Van Santvliet is a Product Manager at Tangent Works. She focuses on bringing TIM’s capabilities to business users, by exposing the underlying functionality through various platforms and tools. This includes Tangent Works’ own web interface TIM Studio, as well as a range of data-related products such as Alteryx, Power BI and Qlik Sense.
Elke is passionate about data in all its aspects, and is always open to discuss the newest trends in AI or dive deep into a specific data science use case.
Tangent Works. (2018, December 11). Tangent Works. Retrieved March 5, 2020, from https://www.tangent.works/
Wikipedia contributors. (2019a, November 10). Structural change. Retrieved March 4, 2020, from https://en.wikipedia.org/wiki/Structural_change
Wikipedia contributors. (2019b, December 12). Structural break. Retrieved March 4, 2020, from https://en.wikipedia.org/wiki/Structural_break