Intelligent Machines in Economics - A macroeconomic application of the LSTM neural network
Author
Schier, Rasmus
Term
4. term
Education
Publication year
2018
Submitted on
2018-05-30
Pages
71
Abstract
Specialet undersøger, om en Long Short-Term Memory (LSTM) neuralt netværk kan forbedre makroøkonomiske prognoser ved at forudsige dansk BNP-vækst ud fra 19 almindelige indikatorer anvendt i nowcasting-litteraturen. Baggrunden er, at makrodata ofte har mange variable men få observationer, og at traditionelle metoder mister effektivitet i meget store datasæt, hvilket gør fleksible maskinlæringsmetoder interessante. Metodisk sammenlignes LSTM’en med en benchmark i form af en principal component regression (PCR) med fem latente faktorer. Begge modeller trænes på en træningsdel af data og evalueres out-of-sample på en testdel efter standard maskinlæringspraksis, målt ved kvadreret fejl, RMSE og MAE, samt med en Diebold–Mariano test for forskelle i prognosekvalitet. Resultaterne viser, at LSTM’en opnår lavere fejlmål end PCR i den opstillede model, men forskellen er ikke statistisk signifikant. LSTM’en kan således anvendes i en makroøkonomisk ramme og kan muligvis forbedres med mere data, men den begrænsede strukturelle fortolkelighed svækker tilliden til prognoserne og giver ingen klar fordel over faktorbaserede modeller i denne opsætning. Specialet peger på fremtidige muligheder for bedre præstation via variational auto-encoders til komprimeret repræsentation af input og større gennemsigtighed gennem lokale forklaringsværktøjer som LIME, men understreger behovet for yderligere forskning i makroøkonomiske regressionsopgaver.
The thesis examines whether a Long Short-Term Memory (LSTM) neural network can improve macroeconomic forecasting by predicting Danish GDP growth from 19 commonly used nowcasting indicators. The motivation is that macro datasets often feature many variables but few observations, and traditional methods can lose efficiency with very large datasets, making flexible machine learning approaches attractive. Methodologically, the LSTM is benchmarked against a principal component regression (PCR) factor model with five latent factors. Both models are trained on a training split and evaluated out-of-sample on a test split following standard machine learning practice, using squared error, RMSE, and MAE, and a Diebold–Mariano test to assess differences in forecast accuracy. Results show that the LSTM achieves lower error measures than the PCR in the presented setup, but the difference is not statistically significant. The LSTM is therefore feasible in a macroeconomic context and may benefit from more data, yet limited structural interpretability reduces trust in its forecasts and yields no clear advantage over factor-based models in this application. The thesis highlights future opportunities via variational auto-encoders for compact input representations and local explanation tools such as LIME to enhance transparency, while noting the need for further research in macroeconomic regression settings.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents