Merge branch 'conclusion-section' into develop

paper.tex

@@ -37,6 +37,9 @@
 \input{tex/4_stu/6_fams}
 \input{tex/4_stu/7_pixels_intervals}
 \input{tex/5_con/1_intro}
+\input{tex/5_con/2_generalizability}
+\input{tex/5_con/3_implications}
+\input{tex/5_con/4_further_research}
 \newpage
 
 \input{tex/glossary}

tex/4_stu/4_overall.tex

@@ -44,13 +44,6 @@ As the non-seasonal \textit{hses} reaches a similar accuracy as its
 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.
-For longer horizons, \textit{hets} provides the highest accuracy.
-Thus, to fit a seasonal pattern, longer training horizons are needed.
-While \textit{vsvr} enters the top three, \textit{hets} has the edge as they
-    neither require parameter tuning nor real-time data.
-
 \begin{center}
 \captionof{table}{Top-3 models by training weeks and average demand
                   ($1~\text{km}^2$ pixel size, 60-minute time steps)}
@@ -206,6 +199,13 @@ While \textit{vsvr} enters the top three, \textit{hets} has the edge as they
 \end{tabular}
 \end{center}
 
+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.
+For longer horizons, \textit{hets} provides the highest accuracy.
+Thus, to fit a seasonal pattern, longer training horizons are needed.
+While \textit{vsvr} enters the top three, \textit{hets} has the edge as they
+    neither require parameter tuning nor real-time data.
+
 In summary, except for high demand, simple models trained on horizontal time
     series work best.
 By contrast, high demand (i.e., $25 < \text{ADD} < \infty$) and less than

tex/4_stu/6_fams.tex

@@ -1,19 +1,6 @@
 \subsection{Results by Model Families}
 \label{fams}
 
-Besides the overall results, we provide an in-depth comparison of models
-    within a family.
-Instead of reporting the MASE per model, we rank the models holding the
-    training horizon fixed to make comparison easier.
-Table \ref{t:hori} presents the models trained on horizontal time series.
-In addition to \textit{naive}, we include \textit{fnaive} and \textit{pnaive}
-    already here as more competitive benchmarks.
-The tables in this section report two rankings simultaneously:
-The first number is the rank resulting from lumping the low and medium
-    clusters together, which yields almost the same rankings when analyzed
-    individually.
-The ranks from only high demand pixels are in parentheses if they differ.
-
 \begin{center}
 \captionof{table}{Ranking of benchmark and horizontal models
                   ($1~\text{km}^2$ pixel size, 60-minute time steps):
@@ -47,6 +34,19 @@ The ranks from only high demand pixels are in parentheses if they differ.
 \end{center}
 \
 
+Besides the overall results, we provide an in-depth comparison of models
+    within a family.
+Instead of reporting the MASE per model, we rank the models holding the
+    training horizon fixed to make comparison easier.
+Table \ref{t:hori} presents the models trained on horizontal time series.
+In addition to \textit{naive}, we include \textit{fnaive} and \textit{pnaive}
+    already here as more competitive benchmarks.
+The tables in this section report two rankings simultaneously:
+The first number is the rank resulting from lumping the low and medium
+    clusters together, which yields almost the same rankings when analyzed
+    individually.
+The ranks from only high demand pixels are in parentheses if they differ.
+
 A first insight is that \textit{fnaive} is the best benchmark in all
     scenarios:
 Decomposing flexibly by tuning the $ns$ parameter is worth the computational

tex/5_con/1_intro.tex

@@ -1,2 +1,6 @@
 \section{Conclusion}
 \label{con}
+
+We conclude this paper by elaborating on how the findings are transferable
+    to similar settings, providing some implications for a UDP's
+    managers, and discussing further research opportunities.

tex/5_con/2_generalizability.tex

@@ -0,0 +1,23 @@
+\subsection{Generalizability of the Methology and Findings}
+\label{generalizability}
+
+Whereas forecasting applications are always data-specific, the following
+    aspects generalize to UDPs with ad-hoc transportation services:
+\begin{itemize}
+\item \textbf{Double Seasonality}:
+The double seasonality causes a periodicity $k$ too large to be modeled by
+    classical models, and we adapt the STL method in the \textit{fnaive} model
+    such that it "flexibly" fits a seasonal pattern changing in a non-trivial
+    way over time.
+\item \textbf{Order Sparsity}:
+The intermittent time series resulting from gridification require simple
+    methods like \textit{hsma} or \textit{trivial} that are not as susceptible
+    to noise as more sophisticated ones.
+\item \textbf{Unified CV}:
+A CV unified around a whole day allows evaluating classical statistical and ML
+    methods on the same scale.
+It is agnostic of both the type of the time series and retraining.
+\item \textbf{Error Measure}:
+Analogous to \cite{hyndman2006}, we emphasize the importance of choosing a
+    consistent error measure, and argue for increased use of MASE.
+\end{itemize}

tex/5_con/3_implications.tex

@@ -0,0 +1,58 @@
+\subsection{Managerial Implications}
+\label{implications}
+
+Even though zeitgeist claims that having more data is always better, our study
+    shows this is not the case here:
+First, under certain circumstances, accuracy may go up with shorter training
+    horizons.
+Second, none of the external data sources improves the accuracies.
+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
+    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,
+    with more than six weeks of training material available.
+Lastly, with just \textit{hets}, that exhibits an accuracy comparable to
+    \textit{hsma} for low and medium demand, our industry partner can likely
+    schedule its shifts on an hourly basis one week in advance.
+
+This study gives rise to the following managerial implications.
+First, UDPs can implement readily available forecasting algorithms with limited
+    effort.
+This, however, requires purposeful data collection and preparation by those
+    companies, which, according to our study, is at least equally important as
+    the selection of the forecasting algorithm, as becomes clear from 
+    investigating the impact of the length of the training horizon.
+Second, the benefits of moving from manual forecasting to automated forecasting
+    include being able to pursue a predictive routing strategy and
+    demand-adjusted shift scheduling.
+At the time the case study data was collected, our industry partner did not
+    conduct any forecasting; the only forecasting-related activities were the
+    shift managers scheduling the shifts one week in advance manually in 
+    spreadsheets.
+Thus, selecting the right forecasting algorithm according to the framework
+    proposed in this study becomes a prerequisite to the much needed
+    operational improvements UDPs need to achieve in their quest for
+    profitability.
+In general, many UDPs launched in recent years are venture capital backed
+    start-up companies that almost by definition do not have a strong
+    grounding in operational excellence, and publications such as the ones by
+    Uber are the exception rather than the rule.
+Our paper shows that forecasting the next couple of hours can already be
+    implemented within the first year of a UDP's operations.
+Even if such forecasts could not be exploited by predictive routing (e.g., due
+    to prolonged waiting times at restaurants), they would help monitoring the
+    operations for exceptional events.
+Additionally, the shift planning may be automated saving as much as one shift
+    manager per city.
+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
+    empirical study suggest that a pixel size of $1~\text{km}^2$ and a time step of
+    one hour are ideal.

tex/5_con/4_further_research.tex

@@ -0,0 +1,45 @@
+\subsection{Further Research}
+\label{further_research}
+
+Sub-sections \ref{overall_results} and \ref{fams} present the models' average
+    performance.
+We did not research what is the best model in a given pixel on a given day.
+To answer this, a study finding an optimal number of outer validation days is
+    neccessary.
+With the varying effect of the training horizon, this model selection is a
+    two-dimensional grid search that is prone to overfitting due to the high
+    noise in low count data.
+Except heuristics relating the ADD to the training horizon, we cannot say
+    anything about that based on our study.
+\cite{lemke2010} and \cite{wang2009} show how, for example, a time series'
+    characteristics may be used to select models.
+Thus, we suggest conducting more detailed analyses on how to incorporate model
+    selection into our proposed forecasting system.
+
+Future research should also integrate our forecasting system into a predictive
+    routing application and evaluate its business impact.
+This embeds our research into the vast literature on the VRP.
+Initially introduced by \cite{dantzig1959}, \gls{vrp}s are concerned with
+    finding optimal routes serving customers.
+We refer to \cite{toth2014} for a comprehensive overview.
+The two variants relevant for the UDP case are the dynamic VRP and
+    the pickup and delivery problem (\gls{pdp}).
+A VRP is dynamic if the data to solve a problem only becomes available
+    as the operations are underway.
+\cite{thomas2010}, \cite{pillac2013}, and \cite{psaraftis2016} describe how
+    technological advances, in particular, mobile technologies, have led to a
+    renewed interest in research on dynamic VRPs, and
+    \cite{berbeglia2010} provide a general overview.
+\cite{ichoua2006} and \cite{ferrucci2013} provide solution methods for
+    simulation studies where they assume stochastic customer demand based on
+    historical distributions.
+In both studies, dummy demand nodes are inserted into the VRP instance.
+Forecasts by our system extend this idea naturally as dummy nodes could be
+    derived from point forecasts as well.
+The concrete case of a meal delivering UDP is contained in a recent
+    literature stream started by \cite{ulmer2017} and extended by
+    \cite{reyes2018} and \cite{yildiz2018}: They coin the term meal delivery
+    routing problem (\gls{mdrp}).
+The MDRP is a special case of the dynamic PDP where the defining
+    characteristic is that once a vehicle is scheduled, a modification of the
+    route is inadmissible.

tex/glossary.tex

@@ -11,9 +11,15 @@
 \newglossaryentry{mase}{
     name=MASE, description={Mean Absolute Scaled Error}
 }
+\newglossaryentry{mdrp}{
+    name=MDRP, description={Meal Delivery Routing Proplem}
+}
 \newglossaryentry{ml}{
     name=ML, description={Machine Learning}
 }
+\newglossaryentry{pdp}{
+    name=PDP, description={Pickup and Delivery Problem}
+}
 \newglossaryentry{rf}{
     name=RF,  description={Random Forest}
 }

tex/references.bib

@@ -44,6 +44,18 @@ howpublished = {\url{https://eng.uber.com/forecasting-introduction/}},
 note = {Accessed: 2020-10-01}
 }
 
+@article{berbeglia2010,
+title={Dynamic Pickup and Delivery Problems},
+author={Berbeglia, Gerardo and Cordeau, Jean-Fran{\c{c}}ois
+        and Laporte, Gilbert},
+year={2010},
+journal={European Journal of Operational Research},
+volume={202},
+number={1},
+pages={8--15},
+publisher={Elsevier}
+}
+
 @article{box1962,
 title={Some statistical Aspects of adaptive Optimization and Control},
 author={Box, George and Jenkins, Gwilym},
@@ -167,6 +179,17 @@ year={2016},
 publisher={Springer}
 }
 
+@article{dantzig1959,
+title={The truck dispatching problem},
+author={Dantzig, George and Ramser, John},
+year={1959},
+journal={Management science},
+volume={6},
+number={1},
+pages={80--91},
+publisher={Informs}
+}
+
 @article{de2006,
 title={25 Years of Time Series Forecasting},
 author={De Gooijer, Jan and Hyndman, Rob},
@@ -196,6 +219,18 @@ volume={116},
 pages={242--265}
 }
 
+@article{ferrucci2013,
+title={A pro-active real-time Control Approach for Dynamic Vehicle Routing
+       Problems dealing with the Delivery of urgent Goods},
+author={Ferrucci, Francesco and Bock, Stefan and Gendreau, Michel},
+year={2013},
+journal={European Journal of Operational Research},
+volume={225},
+number={1},
+pages={130--141},
+publisher={Elsevier}
+}
+
 @article{gardner1985,
 title={Forecasting Trends in Time Series},
 author={Gardner, Everette and McKenzie, Ed},
@@ -329,6 +364,18 @@ year={2018},
 publisher={OTexts}
 }
 
+@article{ichoua2006,
+title={Exploiting Knowledge about Future Demands for Real-time Vehicle
+       Dispatching},
+author={Ichoua, Soumia and Gendreau, Michel and Potvin, Jean-Yves},
+year={2006},
+journal={Transportation Science},
+volume={40},
+number={2},
+pages={211--225},
+publisher={INFORMS}
+}
+
 @article{kim2016,
 title={A new Metric of Absolute Percentage Error for Intermittent Demand
        Forecasts},
@@ -362,6 +409,17 @@ howpublished = {\url{https://eng.uber.com/neural-networks/}},
 note = {Accessed: 2020-10-01}
 }
 
+@article{lemke2010,
+title={Meta-Learning for Time Series Forecasting and Forecast Combination},
+author={Lemke, Christiane and Gabrys, Bogdan},
+year={2010},
+journal={Neurocomputing},
+volume={73},
+number={10-12},
+pages={2006--2016},
+publisher={Elsevier}
+}
+
 @article{ma2018,
 title={Using the Gradient Boosting Decision Tree to Improve the Delineation of
        Hourly Rain Areas during the Summer from Advanced Himawari Imager Data},
@@ -430,6 +488,18 @@ number={5},
 pages={311--315}
 }
 
+@article{pillac2013,
+title={A Review of Dynamic Vehicle Routing Problems},
+author={Pillac, Victor and Gendreau, Michel and Gu{\'e}ret, Christelle
+        and Medaglia, Andr{\'e}s L},
+year={2013},
+journal={European Journal of Operational Research},
+volume={225},
+number={1},
+pages={1--11},
+publisher={Elsevier}
+}
+
 @article{prestwich2014,
 title={Mean-based Error Measures for Intermittent Demand Forecasting},
 author={Prestwich, Steven and Rossi, Roberto and Tarim, Armagan
@@ -442,6 +512,25 @@ pages={6782--6791},
 publisher={Taylor \& Francis}
 }
 
+@article{psaraftis2016,
+title={Dynamic Vehicle Routing Problems: Three Decades and Counting},
+author={Psaraftis, Harilaos and Wen, Min and Kontovas, Christos},
+year={2016},
+journal={Networks},
+volume={67},
+number={1},
+pages={3--31},
+publisher={Wiley Online Library}
+}
+
+@article{reyes2018,
+title={The Meal Delivery Routing Problem},
+author={Reyes, Damian and Erera, Alan and Savelsbergh, Martin
+        and Sahasrabudhe, Sagar and O’Neil, Ryan},
+  year={2018},
+journal={Optimization Online}
+}
+
 @incollection{scholkopf1998,
 title={Fast Approximation of Support Vector Kernel Expansions, and an
        Interpretation of Clustering as Approximation in Feature Spaces},
@@ -501,6 +590,30 @@ number={4},
 pages={715--725}
 }
 
+@article{thomas2010,
+title={Dynamic vehicle routing},
+author={Thomas, Barrett W},
+year={2010},
+journal={Wiley Encyclopedia of Operations Research and Management Science},
+publisher={Wiley Online Library}
+}
+
+@book{toth2014,
+title={Vehicle Routing: Problems, Methods, and Applications},
+author={Toth, Paolo and Vigo, Daniele},
+year={2014},
+publisher={SIAM}
+}
+
+@techreport{ulmer2017,
+title={The Restaurant Meal Delivery Problem: Dynamic Pick-up and Delivery with
+       Deadlines and Random Ready Times},
+author={Ulmer, Marlin and Thomas, Barrett and Campbell, Ann Melissa
+        and Woyak, Nicholas},
+year={2017},
+institution={Technical Report}
+}
+
 @article{vapnik1963,
 title={Pattern Recognition using Generalized Portrait Method},
 author={Vapnik, Vladimir and Lerner, A},
@@ -525,6 +638,18 @@ year={2013},
 publisher={Springer}
 }
 
+@article{wang2009,
+title={Rule Induction for Forecasting Method Selection:
+       Meta-learning the Characteristics of Univariate Time Series},
+author={Wang, Xiaozhe and Smith-Miles, Kate and Hyndman, Rob},
+year={2009},
+journal={Neurocomputing},
+volume={72},
+number={10-12},
+pages={2581--2594},
+publisher={Elsevier}
+}
+
 @article{wang2018,
 title={Delivering meals for multiple suppliers: Exclusive or sharing
        logistics service},
@@ -556,3 +681,10 @@ volume={6},
 number={3},
 pages={324--342}
 }
+
+@article{yildiz2018,
+title={Provably High-Quality Solutions for the Meal Delivery Routing Problem},
+author={Yildiz, Baris and Savelsbergh, Martin},
+  year={2018},
+journal={Optimization Online}
+}