Author Archives: Nikos

Forecasting multiple time series with tsintermittent

I uploaded a new version (1.7) of tsintermittent on CRAN. Apart from fixing a couple of minor issues, a new function has been added to help scaling up forecasting. Recently I had a few requests to add a functionality to use data frames with multiple time series as inputs. I have included a new wrapper function data.frc that does exactly that.

Here is an example. Let us first create some data (10 series of 20 observations each) and call the new function:
> y <- simID(10,20)
> yhat <- data.frc(y,"crost")
The output has two parts, yhat$frc.out will is a data frame with the forecasts for all time series, while yhat$out will be a list with the detailed output for each time series.

Currently all forecasting methods in tsintermittent are supported: crost, crost.ma, tsb, sexsm and imapa.

You can also pass other arguments that are relevant to each function that each called, for example:
> yhat <- data.frc(y,"crost",h=15,type="sba",na.rm=TRUE)
These additional options are documented in each forecasting method.

You may have also noticed that there is a new option for all forecasting methods, na.rm, which removes any NA values from each series in the data frame.

Note that if you call crost(y), only the first time series in the data frame will be used, as the individual functions are designed to deal with individual time series. Only the wrapper function will through the complete set.

Hope you find the new function useful! Other minor changes concern the function simID which now outputs the generated series as a data frame, resolving an inconsistency in the output of that function with the rest of the package functions.

Related Posts

Forecasting with multivariate temporal aggregation: The case of promotional modelling

N. Kourentzes and F. Petropoulos, 2015, International Journal of Production Economics, 181: 145-153. http://dx.doi.org/10.1016/j.ijpe.2015.09.011

Demand forecasting is central to decision making and operations in organisations. As the volume of forecasts increases, for example due to an increased product customisation that leads to more SKUs being traded, or a reduction in the length of the forecasting cycle, there is a pressing need for reliable automated forecasting. Conventionally, companies rely on a statistical baseline forecast that captures only past demand patterns, which is subsequently adjusted by human experts to incorporate additional information such as promotions. Although there is evidence that such process adds value to forecasting, it is questionable how much it can scale up, due to the human element. Instead, in the literature it has been proposed to enhance the baseline forecasts with external well-structured information, such as the promotional plan of the company, and let experts focus on the less structured information, thus reducing their workload and allowing them to focus where they can add most value. This change in forecasting support systems requires reliable multivariate forecasting models that can be automated, accurate and robust. This paper proposes an extension of the recently proposed Muliple Aggregation Prediction Algorithm (MAPA), which uses temporal aggregation to improve upon the established exponential smoothing family of methods. MAPA is attractive as it has been found to increase both the accuracy and robustness of exponential smoothing. The extended multivariate MAPA is evaluated against established benchmarks in modelling a number of heavily promoted products and is found to perform well in terms of forecast bias and accuracy. Furthermore, we demonstrate that modelling time series using multiple temporal aggregation levels makes the final forecast robust to model misspecification.

Download paper.

Update: R code for MAPAx is now available!

Related Posts

Forecasting with Temporal Hierarchies

G. Athanasopoulos, R.J. Hyndman, N. Kourentzes and F. Petropoulos, 2015.

This paper introduces the concept of Temporal Hierarchies for time series forecasting. A temporal hierarchy can be constructed for any time series by means of non-overlapping temporal aggregation. Predictions constructed at all aggregation levels are combined with the proposed framework to result in temporally reconciled, accurate and robust forecasts. The implied combination mitigates modelling uncertainty, while the reconciled nature of the forecasts results in a unified prediction that supports aligned decisions at different planning horizons: from short-term operational up to long-term strategic planning. The proposed methodology is independent of forecasting models. It can embed high level managerial forecasts that incorporate complex and unstructured information with lower level statistical forecasts. Our results show that forecasting with temporal hierarchies increases accuracy over conventional forecasting, particularly under increased modelling uncertainty. We discuss organisational implications of the temporally reconciled forecasts using a case study of Accident & Emergency departments.

Download paper.

Related Posts

SAS-IIF Grant to Promote Research on Forecasting

The International Institute of Forecasters (IIF), in collaboration with SAS, offer financial support research on improving forecasting methods and business forecasting practice. TThis year there will be two grants of $5,000 .

The deadline date for applications is September 30, 2015.

This grant was created in 2002 by the IIF, with financial support from the SAS® Institute, in order to promote research on forecasting principles and practice. The fund is divided to support research in the two basic aspects of forecasting: development of theoretical results and new methods and practical applications with real-world comparisons.

You can learn more about this here.

Related Posts

Forecasting Solar Irradiance: True models, trace optimisation and shrinkage

This is a talk I am giving today at Tasmanian School of Business and Economics at University of Tasmania. It connects two different research areas I am currently working on: solar irradiance forecasting and parameter optimisation under model uncertainty.

You can find the presentation here.

Half of the presentation is based on this paper if you are interested in the details.

Related Posts

Forecasting with Temporal Hierarchies

This is a talk that I am giving today at the University of Sydney Business School. This research builds upon MAPA and cross-sectional hierarchical forecasting, in particular optimal combinations. Temporal hierarchies reconcile across time, resulting in accurate short and long-term forecasts that can lead to aligned plans and decisions.

Temporal hierarchies can be used with any time series and forecasting method, even expert judgement, overcoming the limitations of MAPA, but retaining all its strengths.

You can find the presentation here and a working paper here.

R code for temporal hierarchies will be released soon.

Related Posts

True models, trace optimisation and parameter shrinkage

This is a talk that I gave at Monash University, where I am currently visiting. The topic of this research is exploring ways to avoid the assumption that the postulated model we are using is true for the data generating process of the time series we want to forecast. From this starting point we proceed to develop alternative cost functions based on the idea of trace forecasts.

Amongst the various interesting findings of this work, the one that I am most excited about is that we show that the new cost functions are shrinkage estimators of the univariate information, where the amount of shrinkage is controlled by the forecast horizon. In contrast to conventional shrinkage estimators, such as LASSO, we do not need to estimate a shrinkage parameter, as this is drawn from the forecasting problem directly. Furthermore, shrinkage for both AR and MA processes is feasible within this framework.

You can download the talk here.

A working paper of this work will be released here shortly.
I would also like to thank the audience at Monash University for their many interesting comments.

Related Posts

Presentations at EURO 2015

Last week had been very busy at the 27th European Conference on Operational Research. Within the Analytics, Data Science, Data Mining stream there was a specialised stream on Forecasting & Time Series Prediction that was very successfully organised by Sven F. Crone and Aris Syntetos. The talks have been stimulating and of a high level. Given that the conference was hosted in Glasgow it was very easy for several colleagues of mine to attend and therefore I was involved in a large number of presentations.

Here is the list of the various topics that were presented, click on the title to access the presentation slides.

I think that is probably the record number of presentations I will ever be involved in a conference. To mark the occasion I ensured that we would have some photos with all of us together!

euro2015

euro2015b

Related Posts

DIY forecasting: judgment, models and judgmental model selection

F. Petropoulos, N. Kourentzes, K. Nikolopoulos, 2015, 27th European Conference on operational Research, Glasgow.

In this paper we explore how judgment can be used to improve model selection for forecasting. We investigate the performance of various judgmental model selection methodologies against the benchmark statistical one, based on information criteria. Apart from the simple model choice approach, we examine the efficacy of a model build approach, where experts are asked to identify the structural components (trend and seasonality) of the time series. Based on a large sample of almost 700 participants that contributed in a custom-designed laboratory experiment, we evaluate the performance of individuals and groups of experts in terms of selecting the best model and forecasting performance, identifying major improvements. Finally, we examine how to extend statistical model selection to incorporate additional insights from experts.

Download presentation.
This experiment was supported by the Forecasting Society.

Related Posts

Exponential smoothing parameter estimation for complex seasonal forms: the case solar irradiation forecasting

N. Kourentzes, J.R. Trapero, 2015, 27th European Conference on operational Research, Glasgow.

Renewable energy generation has become more important over the years, bringing more sustainable options to the energy mix of countries. Solar power generation is one such option. Although solar energy is attractive, it brings new forecasting challenges. In order to integrate solar energy into the grid it is important to predict the energy supply accurately, which is dependent on solar irradiation. Large forecast errors can lead to significant costs for the operator. In this work we investigate the use of exponential smoothing to produce univariate forecasts for solar irradiation. These type of models have been explored in the past, due to their significant operational advantages as they are simple to deploy and use and do not require costly additional inputs as numerical weather models do. However, the forecasting performance of exponential smoothing has been challenged in the literature. We argue that there are two key reasons for this. First, this is due to the complex seasonal shapes exhibited in solar irradiance data. Second, conventional exponential smoothing parameter estimation is often not capable of identifying good parameters. We explore both issues by investigating the use of alternative optimisation cost functions, relaxing assumptions about the model form, in order to increase short and long term forecasting accuracy. We find that alternative cost functions have substantial benefits in terms of forecasting accuracy, data requirements and computational costs.

Download presentation.

Related Posts