Merge branch 'supervisor-review' into develop

paper.tex

@@ -1,4 +1,4 @@
-\documentclass[preprint,review,12pt]{static/elsarticle}
+\documentclass[preprint,review,12pt,authoryear]{static/elsarticle}
 
 \input{tex/preamble}
 
@@ -42,9 +42,6 @@
 \input{tex/5_con/4_further_research}
 \newpage
 
-\input{tex/glossary}
-\newpage
-
 \appendix
 \newpage
 \input{tex/apx/tabular_ml_models}
@@ -55,6 +52,8 @@
 \newpage
 \input{tex/apx/peak_results}
 \newpage
+\input{tex/apx/glossary}
+\newpage
 
 \bibliographystyle{static/elsarticle-harv}
 \bibliography{tex/references}

tex/1_intro.tex

@@ -11,8 +11,9 @@ A common feature of these platforms is that they do not operate kitchens but
     related processes in simple smartphone apps, and managing the delivery via
     a fleet of either employees or crowd-sourced sub-contractors.
 
-Various kind of urban delivery platforms (UDP) have received attention in
+Various kinds of urban delivery platforms
-    recent scholarly publications.
+    (\gls{udp}; \ref{glossary} provides a glossary with all abbreviations)
+    have received attention in recent scholarly publications.
 \cite{hou2018} look into heuristics to simultaneously optimize courier
     scheduling and routing in general, while \cite{masmoudi2018} do so
     for the popular dial-a-ride problem and \cite{wang2018} investigate
@@ -63,8 +64,8 @@ Moreover, many of the established forecasting methods can only handle a single
     and often low seasonality (i.e., repeated regular pattern), if at all.
 
 In this paper, we develop a rigorous methodology as to how to build and
-    evaluate a robust forecasting system for an urban delivery platform
+    evaluate a robust forecasting system for an UDP
-    (\gls{udp}) that offers ad-hoc point-to-point transportation of any kind.
+    that offers ad-hoc point-to-point transportation of any kind.
 We implement such a system with a broad set of commonly used forecasting
     methods.
 We not only apply established (i.e., "classical") time series methods but also

tex/2_lit/2_class/2_ets.tex

@@ -1,4 +1,4 @@
-\subsubsection{Na\"{i}ve Methods, Moving Averages, and Exponential Smoothing.}
+\subsubsection{Na\"{i}ve Methods, Moving Averages, and Exponential Smoothing}
 \label{ets}
 
 Simple forecasting methods are often employed as a benchmark for more

tex/2_lit/2_class/3_arima.tex

@@ -1,4 +1,4 @@
-\subsubsection{Autoregressive Integrated Moving Averages.}
+\subsubsection{Autoregressive Integrated Moving Averages}
 \label{arima}
 
 \cite{box1962}, \cite{box1968}, and more papers by the same authors in the

tex/2_lit/2_class/4_stl.tex

@@ -1,4 +1,4 @@
-\subsubsection{Seasonal and Trend Decomposition using Loess.}
+\subsubsection{Seasonal and Trend Decomposition using Loess}
 \label{stl}
 
 A time series $y_t$ may exhibit different types of patterns; to fully capture

tex/2_lit/3_ml/2_learning.tex

@@ -1,4 +1,4 @@
-\subsubsection{Supervised Learning.}
+\subsubsection{Supervised Learning}
 \label{learning}
 
 A conceptual difference between classical and ML methods is the format

tex/2_lit/3_ml/3_cv.tex

@@ -1,4 +1,4 @@
-\subsubsection{Cross-Validation.}
+\subsubsection{Cross-Validation}
 \label{cv}
 
 Because ML models are trained by minimizing a loss function $L$, the

tex/2_lit/3_ml/4_rf.tex

@@ -1,4 +1,4 @@
-\subsubsection{Random Forest Regression.}
+\subsubsection{Random Forest Regression}
 \label{rf}
 
 \cite{breiman1984} introduce the classification and regression tree

tex/2_lit/3_ml/5_svm.tex

@@ -1,4 +1,4 @@
-\subsubsection{Support Vector Regression.}
+\subsubsection{Support Vector Regression}
 \label{svm}
 
 \cite{vapnik1963} and \cite{vapnik1964} introduce the so-called support vector

tex/3_mod/3_grid.tex

@@ -9,9 +9,11 @@ Figure \ref{f:grid} shows how the orders' delivery locations are each
     covering the entire service area within a city.
 This gridification step is also applied to the pickup locations separately.
 The lower-left corner is chosen at random.
-\cite{winkenbach2015} apply the same gridification idea and slice an urban
+Applications of this gridification idea to model location-routing problems
-    area to model a location-routing problem, and \cite{singleton2017} portray
+    can be viewed, for example, in \cite{winkenbach2015}, \cite{bergmann2020},
-    it as a standard method in the field of urban analytics.
+    \cite{janjevic2019}, \cite{snoeck2020}, and \cite{janjevic2020}
+    while \cite{singleton2017} portray it as a standard method in the field of
+    urban analytics.
 With increasing pixel sizes, the time series exhibit more order aggregation
     with a possibly stronger demand pattern.
 On the other hand, the larger the pixels, the less valuable become the

tex/3_mod/7_models/2_hori.tex

@@ -1,4 +1,4 @@
-\subsubsection{Horizontal and Whole-day-ahead Forecasts.}
+\subsubsection{Horizontal and Whole-day-ahead Forecasts}
 \label{hori}
 
 The upper-left in Figure \ref{f:inputs} illustrates the simplest way to

tex/3_mod/7_models/3_vert.tex

@@ -1,4 +1,4 @@
-\subsubsection{Vertical and Whole-day-ahead Forecasts without Retraining.}
+\subsubsection{Vertical and Whole-day-ahead Forecasts without Retraining}
 \label{vert}
 
 The upper-right in Figure \ref{f:inputs} shows an alternative way to

tex/3_mod/7_models/4_rt.tex

@@ -1,4 +1,4 @@
-\subsubsection{Vertical and Real-time Forecasts with Retraining.}
+\subsubsection{Vertical and Real-time Forecasts with Retraining}
 \label{rt}
 
 The lower-left in Figure \ref{f:inputs} shows how models trained on vertical

tex/3_mod/7_models/5_ml.tex

@@ -1,4 +1,4 @@
-\subsubsection{Vertical and Real-time Forecasts without Retraining.}
+\subsubsection{Vertical and Real-time Forecasts without Retraining}
 \label{ml_models}
 
 The lower-right in Figure \ref{f:inputs} shows how ML models take

tex/4_stu/4_overall.tex

@@ -13,36 +13,6 @@ We labeled them "no", "low", "medium", and "high" demand pixels with
     increasing ADD, and present the average MASE per cluster.
 The $n$ do not vary significantly across the training horizons, which confirms
     that the platform did not grow area-wise and is indeed in a steady-state.
-We use this table to answer \textbf{Q1} regarding the overall best methods
-    under different ADDs.
-All result tables in the main text report MASEs calculated with all time
-    steps of a day.
-In contrast, \ref{peak_results} shows the same tables with MASEs calculated
-    with time steps within peak times only (i.e., lunch from 12 pm to 2 pm and
-    dinner from 6 pm to 8 pm).
-The differences lie mainly in the decimals of the individual MASE
-    averages while the ranks of the forecasting methods do not change except
-    in rare cases.
-That shows that the presented accuracies are driven by the forecasting methods'
-    accuracies at peak times.
-Intuitively, they all correctly predict zero demand for non-peak times.
- 
-Unsurprisingly, the best model for pixels without demand (i.e.,
-    $0 < \text{ADD} < 2.5$) is \textit{trivial}.
-Whereas \textit{hsma} also adapts well, its performance is worse.
-None of the more sophisticated models reaches a similar accuracy.
-The intuition behind is that \textit{trivial} is the least distorted by the
-    relatively large proportion of noise given the low-count nature of the
-    time series.
-
-For low demand (i.e., $2.5 < \text{ADD} < 10$), there is also a clear
-    best-performing model, namely \textit{hsma}.
-As the non-seasonal \textit{hses} reaches a similar accuracy as its
-    potentially seasonal generalization, the \textit{hets}, we conclude that
-    the seasonal pattern from weekdays is not yet strong enough to be
-    recognized in low demand pixels.
-So, in the absence of seasonality, models that only model a trend part are
-    the least susceptible to the noise.
 
 \begin{center}
 \captionof{table}{Top-3 models by training weeks and average demand
@@ -198,6 +168,38 @@ So, in the absence of seasonality, models that only model a trend part are
 \hline
 \end{tabular}
 \end{center}
+\
+
+We use this table to answer \textbf{Q1} regarding the overall best methods
+    under different ADDs.
+All result tables in the main text report MASEs calculated with all time
+    steps of a day.
+In contrast, \ref{peak_results} shows the same tables with MASEs calculated
+    with time steps within peak times only (i.e., lunch from 12 pm to 2 pm and
+    dinner from 6 pm to 8 pm).
+The differences lie mainly in the decimals of the individual MASE
+    averages while the ranks of the forecasting methods do not change except
+    in rare cases.
+That shows that the presented accuracies are driven by the forecasting methods'
+    accuracies at peak times.
+Intuitively, they all correctly predict zero demand for non-peak times.
+
+Unsurprisingly, the best model for pixels without demand (i.e.,
+    $0 < \text{ADD} < 2.5$) is \textit{trivial}.
+Whereas \textit{hsma} also adapts well, its performance is worse.
+None of the more sophisticated models reaches a similar accuracy.
+The intuition behind is that \textit{trivial} is the least distorted by the
+    relatively large proportion of noise given the low-count nature of the
+    time series.
+
+For low demand (i.e., $2.5 < \text{ADD} < 10$), there is also a clear
+    best-performing model, namely \textit{hsma}.
+As the non-seasonal \textit{hses} reaches a similar accuracy as its
+    potentially seasonal generalization, the \textit{hets}, we conclude that
+    the seasonal pattern from weekdays is not yet strong enough to be
+    recognized in low demand pixels.
+So, in the absence of seasonality, models that only model a trend part are
+    the least susceptible to the noise.
 
 For medium demand (i.e., $10 < \text{ADD} < 25$) and training horizons up to
     six weeks, the best-performing models are the same as for low demand.

tex/5_con/3_implications.tex

@@ -10,8 +10,9 @@ Somewhat surprisingly, despite ML-based methods` popularity in both business
     and academia in recent years, we must conclude that classical forecasting
     methods suffice to reach the best accuracy in our study.
 There is one case where ML-based methods are competitive in our case study:
-    In a high demand pixel, if only about four to six weeks of past data is
+    In a high demand pixel (defined as more than 25 orders per day on average),
-    available, the \textit{vrfr} model outperformed the classical ones.
+    if only about four to six weeks of past data is available,
+    the \textit{vrfr} model outperformed the classical ones.
 So, we recommend trying out ML-based methods in such scenarios.
 In addition, with the \textit{hsma} and \textit{hets} models being the overall
     winners, incorporating real-time data is not beneficial, in particular,
@@ -53,6 +54,8 @@ We emphasize that for the most part, our proposed forecasting system
     is calibrated automatically and no manual work by a data scientist is required.
 The only two parameters where assumptions need to be made are the pixel size
     and the time step.
-While they can only be optimized by the data scientist over time, the results in our
+The results in our empirical study suggest
-    empirical study suggest that a pixel size of $1~\text{km}^2$ and a time step of
+    that a pixel size of $1~\text{km}^2$ and a time step of one hour are ideal,
-    one hour are ideal.
+    which results in the optimal trade-off
+    between signal strength and spatial-temporal resolution.
+Future research may explore adaptive grid-sizing depending on, for instance, demand density.

tex/apx/glossary.tex

@@ -1,3 +1,6 @@
+\section{Glossary}
+\label{glossary}
+
 % Abbreviations for technical terms.
 \newglossaryentry{add}{
     name=ADD, description={Average Daily Demand}
@@ -138,4 +141,4 @@
                             trained on vertical time series}
 }
 
-\printglossaries
+\printglossary[title=]

tex/meta.tex

@@ -3,7 +3,7 @@
 \journal{Transportation Research Part E}
 \title{Real-time Demand Forecasting for an Urban Delivery Platform}
 
-\author[WHU]{Alexander Hess\fnref{emails}}
+\author[WHU]{Alexander Hess\fnref{emails}\fnref{corresponding}}
 \author[WHU]{Stefan Spinler\fnref{emails}}
 \author[MIT]{Matthias Winkenbach\fnref{emails}}
 \address[WHU]{
@@ -21,20 +21,26 @@ Emails:
     mwinkenb@mit.edu
 }
 
+\fntext[corresponding]{
+The corresponding author is Alexander Hess.
+Use the provided email.
+}
+
 \begin{abstract}
 Meal delivery platforms like Uber Eats shape the landscape in cities around the world.
-This paper addresses forecasting demand into the short-term future.
+This paper addresses forecasting demand on a grid into the short-term future,
+    enabling, for example, predictive routing applications.
 We propose an approach incorporating
-    both classical forecasting
+    both classical forecasting and machine learning methods
-    and machine learning methods.
+    and adapt model evaluation and selection to typical demand:
-Model evaluation and selection is adapted to demand typical for such a platform
+        intermittent with a double-seasonal pattern.
-    (i.e., intermittent with a double-seasonal pattern).
+An empirical study shows that
-The results of an empirical study with a European meal delivery service show
+    an exponential smoothing based method trained on past demand data alone
-    that machine learning models become competitive
+        achieves optimal accuracy,
-    once the average daily demand passes a threshold.
+    if at least two months are on record.
-As a main contribution, the paper explains
+With a more limited demand history,
-    how a forecasting system must be set up
+    machine learning is shown
-    to enable predictive routing.
+    to yield more accurate prediction results than classical methods.
 \end{abstract}
 
 \begin{keyword}

tex/references.bib

@@ -56,6 +56,17 @@ pages={8--15},
 publisher={Elsevier}
 }
 
+@article{bergmann2020,
+title={Integrating first-mile pickup and last-mile delivery
+       on shared vehicle routes for efficient urban e-commerce distribution},
+author={Bergmann, Felix M and Wagner, Stephan M and Winkenbach, Matthias},
+year={2020},
+journal={Transportation Research Part B: Methodological},
+volume={131},
+pages={26--62},
+publisher={Elsevier}
+}
+
 @article{box1962,
 title={Some statistical Aspects of adaptive Optimization and Control},
 author={Box, George and Jenkins, Gwilym},
@@ -376,6 +387,26 @@ pages={211--225},
 publisher={INFORMS}
 }
 
+@article{janjevic2019,
+title={Integrating collection-and-delivery points
+       in the strategic design of urban last-mile e-commerce distribution networks},
+author={Janjevic, Milena and Winkenbach, Matthias and Merch{\'a}n, Daniel},
+year={2019},
+journal={Transportation Research Part E: Logistics and Transportation Review},
+volume={131},
+pages={37--67},
+publisher={Elsevier}
+}
+
+@article{janjevic2020,
+title={Designing Multi-tier, Multi-service-level, and Multi-modal
+       Last-Mile Distribution Networks for Omni-Channel Operations},
+author={Janjevic, Milena and Merchan, Daniel and Winkenbach, Matthias},
+year={2020},
+journal={European Journal of Operational Research},
+publisher={Elsevier}
+}
+
 @article{kim2016,
 title={A new Metric of Absolute Percentage Error for Intermittent Demand
        Forecasts},
@@ -559,6 +590,17 @@ number={3},
 pages={199--222}
 }
 
+@article{snoeck2020,
+title={The value of physical distribution flexibility
+       in serving dense and uncertain urban markets},
+author={Snoeck, Andr{\'e} and Winkenbach, Matthias},
+year={2020},
+journal={Transportation Research Part A: Policy and Practice},
+volume={136},
+pages={151--177},
+publisher={Elsevier}
+}
+
 @article{stitson1999,
 title={Support Vector Regression with ANOVA Decomposition Kernels},
 author={Stitson, Mark and Gammerman, Alex and Vapnik, Vladimir