|
on Forecasting |
| By: | Kasun Bandara; Christoph Bergmeir; Slawek Smyl |
| Abstract: | With the advent of Big Data, nowadays in many applications databases containing large quantities of similar time series are available. Forecasting time series in these domains with traditional univariate forecasting procedures leaves great potentials for producing accurate forecasts untapped. Recurrent neural networks, and in particular Long Short-Term Memory (LSTM) networks have proven recently that they are able to outperform state-of-the-art univariate time series forecasting methods in this context, when trained across all available time series. However, if the time series database is heterogeneous accuracy may degenerate, so that on the way towards fully automatic forecasting methods in this space, a notion of similarity between the time series needs to be built into the methods. To this end, we present a prediction model using LSTMs on subgroups of similar time series, which are identified by time series clustering techniques. The proposed methodology is able to consistently outperform the baseline LSTM model, and it achieves competitive results on benchmarking datasets, in particular outperforming all other methods on the CIF2016 dataset. |
| Date: | 2017–10 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:1710.03222&r=for |
| By: | Itai Areili; Yakov Babichenko; Rann Smorodinsky |
| Abstract: | Bayesian experts with a common prior who are exposed to different evidence often make contradictory probabilistic forecasts. An aggregator who receives the forecasts must aggregate them in the best way possible. This may prove to be a challenge whenever the aggregator is not familiar with the prior or the model and evidence available to the experts. We propose a model where experts provide forecasts over a binary state space. We adapt the notion of regret as a means of evaluating schemes that aggregate their forecasts into a single forecast. Our results show that arbitrary correlation between the experts entails high regret, whereas if there are two experts who are Blackwell-ordered (i.e., one of the experts is more informed) or who have conditionally independent evidence, then the regret is surprisingly low. For these latter cases we construct (nearly) optimal aggregation schemes. |
| Date: | 2017–10 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:1710.02838&r=for |
| By: | Chris McDonald (Reserve Bank of New Zealand) |
| Abstract: | Forecasts of non-tradables inflation have been produced using Phillips curves, where capacity pressure and inflation expectations have been the key drivers. The Bank had previously used the survey of 2-year ahead inflation expectations in its Phillips curve. However, from 2014 non-tradables inflation was weaker than the survey and estimates of capacity pressure suggested. Bank research indicated the weakness in non-tradables inflation may have been linked to low past inflation and its impact on pricing behaviour. This note evaluates whether measures of past inflation could have been used to produce forecasts of inflation that would have been more accurate than using surveys of inflation expectations. It does this by comparing forecasts for annual non-tradablesinflation one year ahead. Forecasts are produced using Phillips curves that incorporate measures of past inflation or surveys of inflation expectations, and other information available at the time of each Monetary Policy Statement (MPS). This empirical test aims to determine the approach that captures pricing behaviour best, highlighting which may be best for forecasting going forward. The results show that forecasts constructed using measures of past inflation have been more accurate than using survey measures of inflation expectations, including the 2-year ahead survey measure previously used by the Bank. In addition, forecasts constructed using measures of past inflation would have been significantly more accurate than the Bank’s MPS forecasts since 2009, and only slightly worse than these forecasts before the global financial crisis (GFC). The consistency of forecasts using past-inflation measures reduces the concern that this approach is only accurate when inflation is low, and suggests it may be a reasonable approach to forecasting non-tradables inflation generally. From late 2015, the Bank has assumed that past inflation has affected domestic price-setting behaviour more than previously. As a result, monetary policy has needed to be more stimulatory than would otherwise be the case. This price-setting behaviour is assumed to persist, and is consistent with subdued non-tradables inflation and low nominal wage inflation in 2017. |
| Date: | 2017–04 |
| URL: | http://d.repec.org/n?u=RePEc:nzb:nzbans:2017/04&r=for |
| By: | Daniel Kosiorowski; Dominik Mielczarek; Jerzy P. Rydlewski |
| Abstract: | In this article, a new nonparametric and robust method of forecasting hierarchical functional time series is presented. The method is compared with Hyndman and Shang's method with respect to their unbiasedness, effectiveness, robustness, and computational complexity. Taking into account results of the analytical, simulation and empirical studies, we come to the conclusion that our proposal is superior over the proposal of Hyndman and Shang with respect to some statistical criteria and especially with respect to their robustness and computational complexity. The studied empirical example relates to the management of Internet service divided into four subservices. |
| Date: | 2017–10 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:1710.02669&r=for |