Research

Job Market Paper

Abstract: The article proposes a framework for non-linear state space modeling, based on a Transformer neural network. The input definition, architecture and estimation process are customized for nowcasting dynamic factors from small macroeconomic time series datasets. The Transformer consists of two Encoders, where the first one distills information from data into a chosen number of latent factors, and the second Encoder uses this factor nowcast to minimize the prediction error. Conventional models, such as the Kalman filter, can be inserted as a prior to guide the estimation.  Monte Carlo experiment finds that the Transformer learns to nowcast dynamic factors more accurately than the Kalman filter, when the process deviates from linear Gaussian.



Working papers

Abstract: Self-attention based Transformers have replaced their recurrent predecessors in sequence transduction, owing to their greater performance and computational efficiency. This article customizes a decoder-only Transformer for macroeconomic panel data. The Transformer is trained to classify time periods labelled with systemic banking crisis dates. The model is used to warn about excess accumulation of fragility in the financial network and to help in timing the counter-cyclical policies. Keywords: Transformer networks, pre-training, machine learning, macroprudential analysis, systemic crises. JEL Classification:  C63, E58, E61, G28.


Abstract: To prevent epidemic escalation early on, contact tracing uses costly resources to identify and quarantine the unknown contagious people. Both the prevailing epidemic situation and the effectiveness of tracing are uncertain. I derive a resource allocation rule, where this uncertainty is quantified using Erlang's C queuing model, which is incorporated into a macro-epidemiological model featuring cautious heterogeneous agents. Using this rule to operate contact tracing results in containing the epidemic in most cases, reducing labor costs, and supporting a livelier economy, compared to the alternative where uncertainty is not accounted for in resource allocation. Keywords: Contact tracing, Queuing theory, Erlang C, Resource allocation. JEL Classification: C18, H12, I18, Q54.


Abstract: The aim of this study is to evaluate the calibration of DSGE model KOOMA of the Ministry of Finance with a SVAR model, which is identified with sign restrictions. I compare impulse response functions from the SVAR model, which are found statistically significant and robust to changes in model specifications, to the equivalent impulse response functions from KOOMA. The findings suggest that KOOMA generally produce impulse responses with same signs as the SVAR model, but there are some differences in the magnitudes and persistence of the responses.



Policy work



Media (in Finnish)