UT researchers help the Estonian Tax and Customs Board to forecast Estonian economic activity
Researchers at the Center of IT Impact Studies at University of Tartu created an application that produces short-term economic forecasts 1 to 12 months into the future based on real time tax declarations of all Estonian companies.
According to the Estonian Tax and Customs Board (ETCB) 98% of Estonian tax revenue comes through voluntary payments and only 2% is collected by way of interventions by the tax authorities. These voluntary payments are reported in monthly tax declarations by approximately 166 000 private companies and public entities that are subject to taxation. The ETCB wanted to use real time information contained in these declarations in producing automated forecasts on the whole economy and its subsectors.
A team of University of Tartu researchers and analysists at the Center of IT Impact Studies (CITIS) took up the task to work out the forecasting models and automate the issuing of predictions. “Economic time-series tend to be heavily seasonal and can be forecasted with standard approaches, but the challenge here was to create models that manage to forecast 1 to 12 months into the future the total turnover of companies, the number of employees and export volume for all sectors separately, at all aggregation levels and all that without much human intervention by the analyst” said Andres Võrk, an economist and lecturer in econometrics affiliated with the CITIS team.
UT researchers came up with an application that allows the user, be it the tax authority or any other actor interested in the forecast, to simply select the indicator to be forecasted, the sector and time window and a set of algorithms will fit the models and issue a numeric forecast (see Figure). The application uses an ensemble method that fits a set of very different forecast models to the data and then uses machine learning to arrive at a set of weights that are used to combine the different models into one single economic forecast.
Figure. User interface of the forecasting application.
“Given the need for flexibility and the very different nature of how different sectors behave, as well as that the wider economy changes over time, we had to come up with a modeling solution that is capable to adjust itself to the different sectors and changing economic circumstances on the go” explained Taavi Unt, the author of the weighting algorithm. “Ensemble models usually manage to be more accurate in forecasts that any single model on its own, plus they are by design flexible as the structure of the economy changes” continued Unt, who is also currently working on his PhD in mathematical statistics at UT.
The forecasting application is only one solution UT researchers built for the tax authority and is part of a wider drive towards predictive economics, i.e. using the data produced by economic actors to have a complete overview of the national economy in real time, as well as having fast and accurate predictions of likely changes in the future. Solutions like that have the potential to make economic policy making faster and smarter, allow for real time impact evaluation of policy changes, as well as reduce administrative burden on companies when filing corporate reports.
Additional information Mihkel Solvak, Senior researcher in technology studies mail: email@example.com