To successfully examine the new ideas, these multidisciplinary projects call for expertise in econometrics, statistics (including time series analysis and machine and statistical learning) and economics/finance. The aim is to address, for their parts, various general gaps in knowledge in the state-of-the-art of econometric research, such as:

Lack of decision-making perspective: We aim to tighten the decisionmaking perspective and the final use of forecasts more explicitly to the bases of the method such as the objective/loss function.

Underestimation of the impact of turning points and directional predictive perspective: We put more effort on getting directional predictions correct when things really matter, such as around the turning points in business and financial cycles and financial crises.

Neglected bounded dynamics: The idea is to respect the natural bounds in variables, such as the recent Zero Lower Bound (ZLB) in nominal interest rates, generally implying a need for new nonlinear methods.

Lack of nonlinear and non-Gaussian structural econometric tools: As a part of methodological development to develop nonlinear structural econometric methods, such as the ones related to impulse response analysis, to facilitate economic interpretations.

Text as data and overconfidence: The idea is to measure overconfidence from historical newspaper articles by using so-called deep learning models.

The projects are connected but aiming to answer the above challenges from somewhat different perspectives and with different methods:

New nonlinear econometric methods (Emil Aaltonen Foundation project) A characteristic joint feature in the above challenges is the limited dependence (i.e. the objective of interest is not necessarily continuous/real number). Thus, limited dependent econometric methods (i.e. generalized linear models) in time series context are generally required to address them. A part of this contribution is to integrate classification-based thinking to common modelling and forecasting problems, i.e. to introduce nonlinear methods based on their directional and economic performance instead of the traditional statistical criteria such as the least squares.

Econometric learning (Academy of Finland) In addition to the above contributions related mainly to the conventional type of econometric methods, the general idea in this project is to develop “econometric learning” concept by integrating increasingly popular and up-to-date machine and statistical learning methods with the economic theory and practice. That is to develop learning-based econometric methods addressing the specific needs and characteristics of macroeconomic and financial applications. The econometric learning methods will be largely driven by their use in forecasting purposes, addressing the required renewal of econometric analysis in modern data-rich environments. In addition to forecasting objectives as such, the new methods are partly based on the nexus between econometric forecasting and the final use of forecasts in decision making that has so far been widely overlooked (as mentioned above).

Text as data (Emil Aaltonen Foundation project) Purpose of the project is to utilize machine learning – especially deep learning models – and text data from historical newspapers to form important variables needed in economic research for longer time samples that have usually been available for economists. A special focus will be given to general overconfidence and the possible relationship it has with various economic phenomena such as banks’ risk-taking and further on to financial crises.