Browsing by Author "Mustonen, Anna"

In this thesis, early estimates of turnover of trade published by Statistics Finland are constructed in order to make forecasts more accurate and available earlier than they currently are. The term for making early estimates is called nowcasting, and it comes from words now and forecasting. The idea is to use information which is published early and possibly at higher frequencies to obtain early estimates before the official information is published. This subject is important because policymakers and researchers follow forecasts when making decisions about fiscal and monetary policies or testing economic models. Also central banks, public and private entities and statistical agencies collect a vast amount of economic data every year. The lags of current forecasts can be quite long; the lags of Statistics Finland’s turnover of trade is 45 days and 75 days. The data used in this work for nowcasting is sales inquiry, consumer survey and first registrations of motor vehicles collected by Statistics Finland and retail trade’s confidence indicator published by Confederation of Finnish Industries. The nowcasting target, the turnover of trade, is divided to four activities, which are estimated separately: total trade, wholesale and retail trade and repair of motor vehicles and motorcycles (vehicles), wholesale trade and retail trade. The methods used in this thesis are dynamic factor model, least absolute shrinkage and selection operator, ridge regression and elastic net. These models in question are chosen by the fact that they try to avoid the problem of overparameterization, which is the case with the data used in this thesis. The nowcasts are made using sales inquiry including all sectors, sales inquiry including only trade sector, sales inquiry including consumer survey balance figure about own economic situation now, when nowcasting target is retail trade, and sales inquiry including first registrations of motor vehicles data, when nowcasting target is vehicles. The estimation period starts from January 2010 and ends in December 2016 and the time series are from January 2000 to December 2016. The estimations are made for four time lags: t+5, t+10, t+20 and t+26. That is, the sales inquiry used for nowcasting is tested after accumulating 5, 10, 20 and 26 days from the beginning of each month. Every time lag is tested with the four models used in this thesis. The nowcasting results are compared to actual values of every activity using mean absolute error and a naïve ARIMA-forecasts are used as benchmarks. Before consumer survey data and first registrations are added in the models, their prediction power is tested using ARIMA-forecasts. The results showed that sales inquiry including all sectors gave better estimations than sales inquiry including only trade. Nowcasts made from sales inquiry including all sectors gave good results when estimating retail trade, somewhat good results when predicting total trade and wholesale trade and worse results when nowcasting vehicles. Consumer survey’s question about own economic situation now did not outperform the sales inquiry estimations when the target is retail trade and when forecasted with ARIMA-model one step ahead. When the consumer survey data was added to the nowcasting models, the results did not improve. First registrations of motor vehicles with vehicles series outperformed the sales inquiry nowcasting results, when forecasted with ARIMA-model one step ahead. The results were quite similar and only slightly better than with sales inquiry data only, when first registrations of motor vehicles was added in the nowcasting models. Retail trade’s confidence indicator published by Confederation of Finnish Industries was left out of further examination, because it did not perform as well as consumer survey. Retail trade nowcast results gave mean absolute errors under 1% at time lags t+20 and t+26 using elastic net method. There is not much difference between t+20 and t+26 and the time lag t+20 seems appropriate for making nowcasts. Consumer survey data did not improve these results, and therefore it can be left out from the nowcasting models. Nowcasting results for vehicles gave the highest mean absolute errors and adding the first registrations data in the models did not significantly improve the results. Other data sources could be considered adding to the models. Nowcasting results from total trade and wholesale trade gave 1%-2% mean absolute errors, but they were not examined with additional data sources. Sales inquiry including all sectors seemed appropriate for nowcasting retail trade, but total trade, wholesale trade and vehicles need more examination.