Multivariate Demand Forecasting for Rental Bike Systems Based on an Unobserved Component Model

Round 1

Reviewer 1 Report

Title of the paper: Multivariate Demand Forecasting for Rental Bike Systems based on an Unobserved Component Model

General observations:

The paper provides a multivariate demand forecasting for rental bike sharing systems using an Unobserved Component Model (UCM). The model helps in predicting the monthly rents of the BSS. The model is illustrated using various data from over 2000 bikes, 297 stations, and 21 municipalities, and is employed in model building. UCM has been used in many areas for forecasting purposes.

Some general observations are as follows:

1)       Please refer to line 161: Chapter may be replaced by “section” in the research.

2)       Use UCM has been found in forecasting telephone call demand, predicting monthly traffic volume, forecasting OECD industrial turning points, forecasting economic time series, supply chain forecasting, Forecasting electricity prices and their volatilities, forecasting seasonal rainfall patterns, etc. Some comparisons of USM with other models may be drawn.

3)       Related work seems very narrow:

Please refer to line 161: “This chapter introduces the models and their components (sections 4.1-4.3).  The following may be considered to enhance related work.

Tych, W., Pedregal, D.J., Young, P.C. and Davies, J., 2002. An unobserved component model for multi-rate forecasting of telephone call demand: the design of a forecasting support system. International Journal of forecasting, 18(4), pp.673-695.

Bian, Z., Zhang, Z., Liu, X. and Qin, X., 2019. Unobserved component model for predicting monthly traffic volume. Journal of Transportation Engineering, Part A: Systems, 145(12), p.04019052.

Loidl, M., Witzmann-Müller, U. and Zagel, B., 2019. A spatial framework for Planning station-based bike sharing systems. European Transport Research Review, 11(1), pp.1-12.

4)       How the validity of selected exogenous factors has been verified, as documented in Table 1.

5)       Please refer to lines 270-272: As a result, two more promising sets were found (SF4 & SF5), and  “ Thus, the experiments found .…. model’s forecasting quality.” may be explained further for clarity.

6)       Please refer to line 161: “ The evaluation shows that the forecast quality of the model is mostly high…however, have limited quality.”, the statement may be supported by statics to draw the inference. The result of different errors calculated may be quoted.

7)       Please refer to line number 223: “Corona restrictions in the form of lockdowns, such as the first lockdown (March - June 2020), are also recognizable” and “…. “Such as the number of vacations or corona counts did not significantly affect the model’s forecasting quality” statements seem to be contradictory.

8)       The figure caption looks very long and distracts the reading flow, hence may be transferred to the appropriate place in the manuscript.

9)       Please refer to the heading “3. Data, endogenous and exogenous effects and methods” there is no discussion over endogenous

Author Response

1) Please refer to line 161: Chapter may be replaced by “section” in the research.

The terminology chapter has been replaced by section and is now used consistently.

2) Use UCM has been found in forecasting telephone call demand, predicting monthly traffic volume, forecasting OECD industrial turning points, forecasting economic time series, supply chain forecasting, Forecasting electricity prices and their volatilities, forecasting seasonal rainfall patterns, etc. Some comparisons of USM with other models may be drawn.

A corresponding passage for the application of UCMs has been added to the introduction.

3) Related work seems very narrow: Please refer to line 161: “This chapter introduces the models and their components (sections 4.1-4.3).  The following may be considered to enhance related work.

The related work was supplemented with additional sources and revised. Bian et al. (Unobserved component model for predicting monthly traffic volume) has already been mentioned and described.

4) How the validity of selected exogenous factors has been verified, as documented in Table 1.

Information on the selection of exogenous factors was complemented.

5)  Please refer to lines 270-272: As a result, two more promising sets were found (SF4 & SF5), and  “ Thus, the experiments found .…. model’s forecasting quality.” may be explained further for clarity.

The experiments will now be explained in more detail at this point.

6)  Please refer to line 161: “ The evaluation shows that the forecast quality of the model is mostly high…however, have limited quality.”, the statement may be supported by statics to draw the inference. The result of different errors calculated may be quoted.

The text was extended by the deviating residuals, so that it is clear when which threshold values are exceeded.

7) Please refer to line number 223: “Corona restrictions in the form of lockdowns, such as the first lockdown (March - June 2020), are also recognizable” and “…. “Such as the number of vacations or corona counts did not significantly affect the model’s forecasting quality” statements seem to be contradictory.

The statements may seem contradictory but are not. We also state this in section 6. To highlight this, we now refer to the relevant passage in section 4.

8) The figure caption looks very long and distracts the reading flow, hence may be transferred to the appropriate place in the manuscript.

The captions were chosen deliberately. They record key observations that complement the continuous text. We have reduced the caption length for some figures.

9) Please refer to the heading “3. Data, endogenous and exogenous effects and methods” there is no discussion over endogenous

The definition and use of the endogenous variable is now explained.

Reviewer 2 Report

 

The author has contributed well to this article but needs to summarize the below questions.

1. The author has explained well in the abstract stating that for monthly rents of 2000 bikes, 297 stations are considered, but when relating this in the finding this information has not been utilized to find. Why have these constraints have used and with the help of these constraints what the author has done to be elaborated?

 

2. Author could include the results achievements or summarization in the abstract section with numerical results.

 

3. The key contribution and novelty have not been detailed in the manuscript. The author needs to include it in the introduction section.

 

4. The literature review could be described logically.

 

5. How the author contributed model is better than the other existing model?

 

6. The output part of Fig. 4-6 is to be explained properly with any comparative measures.

 

7.  The author analyzed the existing algorithm, Does the author contribute any new method or technique to obtain this?

Author Response

1) The author has explained well in the abstract stating that for monthly rents of 2000 bikes, 297 stations are considered, but when relating this in the finding this information has not been utilized to find. Why have these constraints have used and with the help of these constraints what the author has done to be elaborated?

This information about the bike sharing system is not a restriction, but is intended to inform the reader about the service and its scale.

2) Author could include the results achievements or summarization in the abstract section with numerical results.

We have added corresponding numerical results in terms of model quality and error metrics to the abstract.

3) The key contribution and novelty have not been detailed in the manuscript. The author needs to include it in the introduction section.

The introduction and the paper contributions have been updated in detail.

4) The literature review could be described logically.

The Related Work section has been restructured and more sources have been added.

5) How the author contributed model is better than the other existing model?

Our study shows that the addition of user/system related exogenous variables increase the forecasting quality. Moreover, we show that a mixed set of exogenous factors from different categories performs better. This separates the model from previously existing models.

6) The output part of Fig. 4-6 is to be explained properly with any comparative measures.

For the evaluation, classical metrics for model fitness such as AIC/BIC and error metrics such as RMSE and MASE were used. In addition, we have defined three quality levels for residuals to assess the forecast performance for individual months more precisely. We have clarified the text in certain places.

7) The author analyzed the existing algorithm, Does the author contribute any new method or technique to obtain this?

Our focus in this study was not to introduce new techniques for general forecasting of time series. Our contribution is the application and development of a UCM for the first time for bike sharing systems and to evaluate the existing multivariate models. New is the addition of further exogenous factors and the experiments performed as described in the paper.

Round 2

Reviewer 2 Report

The author has done the revision well but still needs to be clarified more on the below comments.

1. The comments given for output part of Fig. 4-6 is to be explained properly with any comparative measures is not clear can the author disclose where it has mentioned in the article.

2. Even though this model has not been existed and a new one, can author elaborate how far the outcome can be obtained/ 

Author Response

1. The comments given for output part of Fig. 4-6 is to be explained properly with any comparative measures is not clear can the author disclose where it has mentioned in the article.

The forecast is of low quality if the threshold of 15,000 for the difference and the 50% interval are exceeded. The threshold values for the evaluation of the residuals were chosen arbitrarily and are just an example. Please note that an evaluation of forecast quality in other scientific fields or applications might demand other thresholds or more elaborate statistical methods. We have now clarified this in the paper.


2. Even though this model has not been existed and a new one, can author elaborate how far the outcome can be obtained/ 

We are sorry. However, we don´t understand the question. The model is reported in detail and the data can be handed out upon request.