Lancaster CMAF Friday Forecasting Talks

The abstract: Andy McCann talks about his experience of learning forecasting models and applying them to forecast daily Accident and Emergency attendances and other healthcare data.  Initially using a Dynamic Harmonic Regression model with ARIMA errors and covariates including school holidays, with more recent work using the Facebook Prophet library.

Speaker: Andrew McCann

Bio: Andy provides data analysis, modelling and visualisation services within the MLCSU Clinical Team in many areas, including monitoring pressure in Emergency Departments, forecasting attendances, predicting seasonal bed demand, Length of Stay initiatives and identifying and alerting local systems to emerging Covid pressures. With a degree from Cambridge University in Economics with statistics options, Andy has a solid background in statistical techniques. Before joining the NHS in September 2018, Andy was the Head of Data Analytics for a large PLC and previously ran his own economic research consultancy.

The abstract: In this webinar, John and Aris will talk about their book, “Intermittent Demand Forecasting. Context, Methods and Applications”. They will explain why the topic of the book is important in business and how some typical inventory – forecasting problems can be solved using state of the art robust approaches.

Speaker 1: John Boylan

Bio: John is Professor of Business Analytics at Lancaster University. His research interests are focused on supply-chain forecasting and include intermittence, seasonality and information sharing. He is an Editor-in-Chief of the Journal of the Operational Research Society and President of the International Society for Inventory Research.

Speaker 2: Aris Syntetos

Bio: Aris is Professor of Operational Research at Cardiff University, where he researches forecasting for improving inventory, manufacturing, and supply chain decisions. He holds the DSV Chair of Logistics and Manufacturing, and he is the Editor-in-Chief of the IMA Journal of Management Mathematics. 

Speaker: Anne-Flore Elard

The abstract: “If dataset A can be ingested on top of the data already used by the Machine Learning models, the forecast results would be 10% more accurate right?. Then if I add dataset B, it will add 5% additional accuracy so with the whole data we should get 95% accuracy.” “Why would I need to spend time on basic statistical forecast when I can go straight to Machine Learning models?” Here are two sample quotes I hear almost every day rooted in the concept that Machine Learning models applied to forecasting are superior to statistical forecasting models in that they are intrinsically more accurate. In this perhaps iconoclastic discussion, I will explore the value-add of Machine Learning for industry-based forecasts. First, let’s look at some research and review the accuracy obtained from different methods in the case of highly random data or specific patterns. Second, as a practitioner, I will spend time discussing the cost component of Machine Learning implementations, both quantitatively and qualitatively with interpretability. Last but not least, I will highlight some lessons learnt over the years on how to effectively implement Machine Learning methods to industry-focused forecasts.

Bio: Anne-Flore Elard is a practitioner in data, data science and analytics, trained in Statistics and Management with an MBA from MIT Sloan. Driven by the value-add that well-managed data and data science can bring to businesses, she has led ML/AI services teams at Deloitte, Scotiabank and Kinaxis. Today, she works at Kinaxis on a portfolio of advanced analytics innovations with the goal to bring them to the market.

Discussant: Nikos Kourentzes

Speaker: Thanos Goltsos

The abstract: Environmental sustainability is a global priority as finite resources are being depleted at an unprecedented rate. The traditional linear economic model of make–use–dispose is leading to irreparable ecological damage and is no longer a viable option, and the shift towards a Circular Economy (CE) and a make–use–reuse model is already underway. The transition to a CE is a complex undertaking that has major implications for operations management. Circular strategies and operations such as repair, remanufacturing, repurposing, and others, must now deal with extra uncertainties stemming from the core (used product) acquisition and transformation process. These operations rely upon accurate forecasts of demand, returns and net demand (demand minus returns) requirements. Net demand (forecasting) departs from traditional demand (forecasting) on two fronts: a) its returns (forecast) constituent depends on past (and at times within-lead time) demand, and b) both net demand and its forecasts can take positive, negative, or zero values. These issues have made net demand characterisation and forecasting challenging, and the measurement of its accuracy problematic. In this upcoming webinar I will discuss concepts, methods, and challenges on the intersection of three areas: circular economy and strategies therein, forecasting, and forecast accuracy measurement.

Bio: Thanos holds an MEng in Mechanical Engineering (incl. MSc in Industrial Management) from Aristotle University of Thessaloniki, an MSc in Business Analysis and Management (hons) from Loughborough University, and a PhD in operations management from the Management Science Laboratory of Athens University of Business and Economics. His research and teaching interests revolve around forecasting and inventory control with a focus on the operations management side of the Circular Economy,  (forward and) reverse supply chains, simulation, and optimisation.

Discussant: Robert Fildes

The abstract: Academics and practitioners working in the area of forecasting are used to applying their knowledge to variety of areas to solve many theoretical and practical problems. However, is forecasting a solution to everything, or are there any cases, when forecasters should abstain from forecasting? Paul Goodwin and Stephan Kolassa will discuss this question in the Lancaster CMAF webinar, expressing their views on the subject, debating and then answering questions of the audience.

Speaker 1: Paul Goodwin

Bio: Paul is Emeritus Professor of Management Science at the University of Bath. He has a PhD from Lancaster University, and his research is primarily concerned with the role of judgment in forecasting and decision making. He is a Fellow of the International Institute of Forecasters and has acted as a consultant for a range of organizations. Until 2015 he was an Editor of the International Journal of Forecasting and, since 2004, he has written the Hot New Research column in Foresight: The International Journal of Applied Forecasting. His recent books include Decision Analysis for Management Judgment 5th edition (co-authored) (Wiley), Forewarned: A Sceptic’s Guide to Prediction (Biteback), and Something Doesn’t Add Up (Profile Books).

Speaker 2: Stephan Kolassa

Bio: Stephan is a Data Science Expert at SAP Switzerland. His responsibilities include the algorithmic, statistical and forecasting aspects of SAP’s retail offerings, from user research across prototyping to training. He also does some academic research on the side, serving as an Honorary Researcher at the Centre for Marketing Analytics and Forecasting at Lancaster University Management School, and as Associate Editor for Foresight: The International Journal of Applied Forecasting.

Speaker: Anna Sroginis

The abstract: With the increase in data sources and frequency of decision-making organisations see an increase in the volume of forecasts that needs to be generated. Nonetheless, several studies verify that human intervention in forecasting remains a common practice. There are several ways that experts can augment statistical forecasts with judgment: (i) adjusting forecasts individually for a single item; (ii) batch-adjusting: correcting several time series or categories at the same time; (iii) model tuning, indicating a location of corrections rather than size and feeding it to a statistical model, for example, by introducing indicator variables in a regression model. The literature has explored extensively the first category, but much less the other two. Yet, these are easily scalable for many products at once, making it easier and faster for forecasters to implement changes. There is limited research in the effectiveness and performance of these approaches. Furthermore, due to their ease of use, both batch-adjustments and model tuning might be overused and, as a result, potentially lose their effectiveness. For instance, in model tuning, introducing indicators for spurious events may result in overfitting rather than augmenting the statistical models which increasingly employ more sophisticated algorithms. Using a case study from a UK retailer, which exhibits all three behaviours of adjustments, we provide empirical evidence of the efficacy of these alternatives, as well as explore the conditions where each alternative may be preferable.

Bio: Anna Sroginis is a lecturer in Management Science at Lancaster University, UK. Her research focuses on judgment in forecasting. In particular, the intersection between human experts and support systems with their algorithms is crucial to make better decisions in practice.

Discussant: John Boylan

Speaker: Evangelos Spiliotis

The abstract: Machine Learning (ML) methods, such as Neural Networks and Regression Trees, have been proven to be very accurate solutions for time series forecasting. The results of many recent forecasting competitions and studies strongly support their use and, therefore, companies and organisations have increased their expectations in terms of performance. Nevertheless, the post-sample accuracy of ML methods is significantly affected by the hyper-parameters used and the explanatory variables selected for making forecasts, especially when the data available for training are noisy or limited in number. If not tuned properly, ML methods may overfit historical observations and adapt poorly to new, previously unseen data. In this webinar, Evangelos from National Technical University of Athens will talk about two techniques that can be used to make ML methods generalise better, namely cross-validation and data augmentation. These techniques have been employed by the winners of several time series forecasting competitions, including the M5, highlighting their potential and practical importance.

Bio: Evangelos Spiliotis is a Research Fellow at the Forecasting & Strategy Unit, National Technical University of Athens (NTUA), where he also serves as Coordinator. He graduated from the School of Electrical and Computer Engineering, NTUA, where he also got his PhD in Forecasting Support Systems. His research interests include time series forecasting, decision support systems, machine learning, optimisation, and energy efficiency and conservation. He co-organised the M4 and M5 forecasting competitions.

Discussant: Sven Crone

Speaker: Bahman Rostami-Tabar

The abstract: When forecasts are required over the lead-time period, forecasters are presented with three distinct time series to select from: i) the original, ii) the non-overlapping and iii) the overlapping temporally aggregated time series. For the former, the forecast is first generated for h-step ahead and then aggregated to get the total value over the lead-time, however for the latter cases, the time series is first aggregated to match the lead-time period and then 1-atep ahead forecast is produced. Very often, practitioners are encouraged to aggregate the time series to a frequency relevant to the decisions the eventual forecasts will support, using non-overlapping temporal aggregation. Using M4 competition data, we design and execute a full factorial experiment exploring the forecast accuracy of three approaches by varying the lead-time. We then develop an approach to combine forecasts generated from the three-time series that results in more accurate forecast. Moreover, we investigate how temporal aggregation affects time series features and examine the association between the original time series features and the performance of approaches using machine learning

Bio: Bahman is Associate Professor in Management Science at Cardiff Business School, Cardiff University, UK. Bahman holds a Ph.D. in Industrial Engineering from the University of Bordeaux, France. He is the founder of Forecasting for Social Good and Democratising Forecasting initiatives sponsored by the International Institute of Forecasters (You can check out these initiatives at https://www.f4sg.org/). In his research, he has been developing and using forecasting models in healthcare, humanitarian operations and supply chain. He has been working with many organisations including the National Health Service (NHS), Welsh Ambulance Service Trusts (WAST), United States Agency for International Developments (USAID) and John Snow Inc.

Discussant: Juan Ramón Trapero Arenas

Speaker: Huijing Chen

The abstract: In the context of short-term demand forecasting, businesses are often required to forecast the demand of seasonal products, based on very few complete seasonal cycles of data. This makes the forecasting task difficult, but one possible solution is to make use of the readily available cross-sectional information from a homogeneous group. We propose a framework based on vector ETS (VETS), and provide a conceptual contribution by devising a taxonomy of pure additive and multiplicative ETS models, bringing together the factors of parameters, initial values and components (PIC). This framework is general in the sense that it is not limited to, although inspired from, the modelling of seasonal components. It can be applied to level, trend (including damped trend) and any possible combination of these features. Insights drawn from this research on the impact from improved efficiency and flexibility of estimation on multivariate forecasting will be discussed.

Bio: Dr Huijing Chen is a Senior Lecturer in Operations Management and holds a PhD in short-term seasonal demand forecasting. She has published in high quality journals such as European Journal of Operational Research, International Journal of Forecasting and Journal of the Operational Research Society. She has won an EPSRC-funded research project examining hierarchical seasonal demand forecasting, and recently a knowledge transfer project (KTP) from Innovate UK to investigate the potential of preventing and reducing fresh food waste through data analytic

Speaker: Johann Robette

The abstract: Like any other business activity, forecasting must monitor, demonstrate and defend its performance. What is its actual added value to the business?

In this webinar, we will share some empirical results obtained using the M5 competition data set. We will discuss the following points:

• Why is assessing the real added value of a forecast so important?
• Does improving the accuracy of a forecast necessarily lead to better decisions?
• How can the quality of a forecast be assessed from a business perspective?
• Where is it worthwhile to improve a forecast and where is it a waste of time and energy?
• What exciting new perspectives do cost-oriented metrics and digital twins open up?

In this presentation Johann discussed the surprising relationship between forecast accuracy and value!

Bio: Johann is Customer Success Manager at Vekia, an innovative French-based SaaS software company he joined in 2009. Passionate about Supply Chain optimisation, Johann has also developed a strong interest in business issues (graduated from Stanford SPCD in 2015). As such, he has taken on various business-oriented roles at Vekia and now looks after a portfolio of key accounts, including Fortune 500 companies. In addition, he actively supports supply chain courses in several French universities and volunteers in an educational and charitable organisation.

Discussant: Prof. John Boylan

Speaker: Kai Hoberg

The abstract: Internet-of-Things (IoT) systems that monitor inventories and sales are the next technological advancement in demand forecasting and replenishment. Unlike traditional systems that record purchases at the point-of-sale, these novel systems can track product usage via smart connected devices at the point-of-consumption, i.e., directly at the end user. This usage data promises to be a valuable basis for automated ordering and replenishment processes. We study such a system in the context of professional coffee machines and use a large data set from a leading service provider to develop models for demand forecasting, inventory management and handling inventory inaccuracy.

Bio: Kai Hoberg is Head of Department of Operations and Technology and Professor of Supply Chain and Operations Strategy at Kühne Logistics University in Hamburg, Germany. He was Assistant Professor of Supply Chain Management at the University of Cologne and received his PhD at Münster University. Before returning to academia, Kai worked as a strategy consultant and project manager in the operations team of Booz & Company. He was a visiting scholar at universities such as Cornell University, Israel Institute of Technology, University of Singapore, or University of Oxford. His current research focuses on improving supply chain planning using the latest SC technologies.

Discussant: Anna-Lena Sachs

Speaker: Shari de Baets

The abstract: Forecasting algorithms provide a significant opportunity for forecasters to improve their accuracy. However, once people have seen an algorithm err, they are quick to abandon it. This phenomenon is known as ‘algorithm aversion’ (Dietvorst, Simmons, & Massey, 2015). However, recently, Logg, Minson, & Moore (2019) have reported the opposite effect: results from six experiments showed that people preferred the advice of an algorithm over that of a person. The authors termed this ‘algorithm appreciation’. How can these findings be reconciled? The effects of feedback and visual aids versus labels are explored and discussed.

Bio: Dr. Shari De Baets is a senior post-doctoral researcher from Belgium. She has a Msc in Industrial Pschology (KULeuven, Belgium), a PhD in Applied Economics (Ghent University, Belgium & University College London, UK) and currently works at the Operations department of the Faculty of Economics and Business Administration of Ghent University (Belgium). Her work is multidisciplinary and combines lessons from behavioural economics and decision-making with knowledge on forecasting and project management. She is the chair of the Early Career Researchers’ section of the International Institute of Forecasting, and the President of European Operations and Chief Forecasting Scientist of the Institute for Neuro & Behavioral Project Management.

Discussant: Paul Goodwin

Speakers: Chris Regan & Dmitrii Ishutin

The abstract: If you live in the UK, you have probably seen or heard a lot about rising electricity and gas tariffs due to unprecedented volatility in the energy market. Some even compared a surge in gas prices with a rise of Bitcoin. Indeed, the energy market is becoming increasingly volatile, especially as the worldwide economy transitions to Net Zero. Brady develops software that helps energy traders navigate through this volatility wave – protect their positions, reduce risks, and achieve P&L. In this webinar, Chris and Dmitrii from Brady will talk about two predictive models they developed for their customers – price direction prediction and liquidity prediction models. Tune in to learn more about the energy markets!

Bio:

Chris Regan is the Product Director at Brady. He is also engaged in other ventures in the energy industry as the Founder of Energy Trading Advisory, Bain Special Advisor for Power Trading, Former Chair of Energy UK Wholesale Markets sub-committee and Head of Power Trading SWGT. At Brady, Chris is leading on the delivery of a short-term power trading SaaS platform. Brady’s PowerDesk solution will allow energy traders to operate in this new decarbonised, intermittent and decentralised world of energy trading. Chris has a background in Physics and complements this with an MBA from INSEAD.

Dmitrii Ishutin is a Quantitative Analyst at Brady. Dmitrii leads development of the advanced analytics module of Brady’s SaaS product for short-term energy traders – PowerDesk Edge. Dmitrii has over 5 years of energy industry experience. Before Brady, Dmitrii worked at EDF Energy (the largest energy company in the UK) as a trader on the curve trading desk where he managed long-term power and carbon hedging for both consumption and generation arms of EDF. Before that, he was a senior data scientist in the volume forecasting team.

Discussant: Jethro Browell

Speaker: Anna-Lena Sachs

The abstract: Inventory decisions are often subject to behavioural biases. We consider a practical setting where a manufacturer makes inventory decisions for several perishable products on a daily basis. We analyse the manufacturer’s decisions and find that several biases that were previously found in laboratory experiments are also present in this setting. We also find a new, group aggregation bias which can be attributed to the multi-product nature of the problem. Even though it would be optimal to determine an individual optimal service level per product, the manufacturer simplifies the problem by optimising service levels per product groups. We find that the manufacturer achieves the target service level effectively, but not efficiently. We provide rationales for the manufacturer’s behaviour, discuss the potential financial gains from de-biasing inventory decisions and provide managerial implications for companies facing inventory decisions.

Bio: Anna-Lena Sachs is a Lecturer in Predictive Analytics at Lancaster University and a member of CMAF. Her research focuses on inventory management, the intersection to forecasting and behavioural operations management. With her research, she has informed the practices at companies in different industries, in particular, in retailing, transportation and healthcare. She has developed several quantitative models to support decision makers and make better inventory decisions. She conducts laboratory and field experiments to analyse how human decision makers actually behave and what type of decision support is useful in practice.

Discussant: Stephan Kolassa

Speaker: Nicolas Vandeput

The abstract: How to improve forecasting accuracy in supply chains? In this webinar, we will discuss which best practices to apply when forecasting demand in supply chains.

Here is the step-by-step guide to demand forecasting excellence:

1. Pick the right granularity and horizon
2. Collect demand (and not sales)
3. Use meaningful forecasting metrics
4. Track Forecast Value Added

In this webinar, Nicolas will discuss these steps and their implications.

Bio: Nicolas Vandeput is a supply chain data scientist specialized in demand forecasting and inventory optimization. He founded his consultancy company SupChains in 2016 and co-founded SKU Science — an online demand forecasting platform — in 2018. Passionate about education, Nicolas is both an avid learner and enjoys teaching at universities: he has taught forecasting and inventory optimization to master students since 2014 in Brussels, Belgium. Since 2020 he is also teaching both subjects at CentraleSupelec, Paris, France. He published Data Science for Supply Chain Forecasting in 2018 (2nd edition in 2021) and Inventory Optimization: Models and Simulations in 2020.

Discussant: Prof. John Boylan

Speaker: Prof. Konstantinos Nikolopoulos

The abstract: We provide forecasts on how the Kingdom of Saudi Arabia can reduce its dependency on the oil sector. This is a very timely quest, given the negative prices of oil in the peak of the COVID-19 pandemic. The forecasting task involves estimating the contribution of the top-5 sectors of the GDP in 20-years-time.  The study involves 4 sequential experiments and 110 participants with increasing levels of expertise: novices (19), semi-experts (73), and experts (18). The first two experiments involved forecasting individually with Unaided Judgment and Structured Analogies; the third formed Interaction Groups. In these experiments, participants did not know for which country they produced forecasts; in the final experiment, we revealed the country name and rerun the third experiment. We demonstrate that the proposed forecasting framework can comprehensively identify areas of long-term GDP diversification, and inform policymaking; furthermore, it can quantify the contribution of each GDP sector to the economy.

Bio: Dr. Konstantinos (Kostas) Nikolopoulos is the Professor in Business Information Systems and Analytics at Durham University Business School. He is an Associate Editor of Oxford IMA "Journal of Management Mathematics" and the "Supply Chain Forum, an International Journal" (Taylor & Francis); he is also the Section Editor-In-Chief for the "Forecasting in Economics and Management" section in the MDPI open access journal "Forecasting". Konstantinos’ work has been consistently appearing in the International Journal of Forecasting (29 outputs) but also in journals for broader audiences including the Journal of Operations Management, the European Journal of Operational Research, and the Journal of Computer Information Systems.

Discussant: Robert Fildes

Speaker: Rebecca Killick

The abstract: In many organisations, accurate forecasts are essential for making adequate informed decisions for a variety of applications from inventory management to staffing optimization. Whatever forecasting model is used, changes in the underlying process can lead to inaccurate forecasts, which will be damaging for decision making.  At the same time forecasting models are becoming increasingly complex and identifying change through direct modelling is problematic. We present a novel framework for monitoring forecasts to ensure they remain accurate. By utilizing sequential changepoint techniques on the forecast errors, our framework allows for the real-time identification of potential changes in the process caused by various external factors. We demonstrate the effectiveness of this framework on numerous forecasting frameworks through simulations, and show its effectiveness over alternative approaches. Finally, we present two concrete examples, one from Royal Mail parcel delivery volumes and one from NHS A&E admissions relating to gallstones.

Bio: Dr Killick is Associate Professor in Statistics at Lancaster University, UK. Her primary research interests lie in the development of novel methodology for the analysis of univariate and multivariate nonstationary time series models; changepoints and locally stationary models. This covers many topics including developing models, model selection, efficient estimation, diagnostics, clustering and prediction.  Rebecca is highly motivated by real world problems and has worked with data in a range of fields including Bioinformatics, Energy, Engineering, Environment, Finance, Health, Linguistics and Official Statistics.

Discussant: Prof. Nikolaos Kourentzes

Related R package: https://github.com/grundy95/changepoint.forecast

Speaker: Sven Crone

The abstract: Artificial Intelligence and Machine Learning (AI/ML) reign high on the Gartner hype cycle, promising new business models, and capturing the imagination of many industry executives well outside the digital industries of Facebook, Google and Uber. However, despite dozens of Proof-of-Concept studies in demand planning, AI/ML is struggling to prove significant increases in planning accuracy. One limiting factor, we will argue, is that the underlying data for demand planning is fundamentally limited. After a very brief introduction into AI/ML we will explore the fundamental difference in data between image/voice/video recognition and typical pharmaceutical demand planning, where we data size and data labels are often very sparse. As this limits the benefit of the AI/ML algorithms of Google & co, we will show a case study where we have enhanced limited datasets in demand panning and thus showed promise in using AI/ML for Janssen (a Johnson & Johnson company). We will close by showing more favourable data conditions and successful AIML forecasts in container shipping for Hapag-Lloyd, forecasting consumer demand for FMCG manufacturer Beiersdorf, and beverage forecasting for beer Manufacturer Anheuser-Busch InBev, which have been forecasting with AI for up to 10 years.

Bio: Sven is an Assistant Professor in Management Science at Lancaster University, UK, where his 15+ years in research on forecasting algorithms and applications of AI/ML led to 80+ publications and ranks him in the top 30 in Europe (citations by Google Scholar). Sven regularly provides in-house training courses on forecasting and analytics for companies, and is a regular conference chair and keynote speaker at international conferences in forecasting, analytics, and S&OP/IBP. He is also CEO and founder of iqast.de, a university spin-off pioneering AIML algorithms in software that work as an add-on to the dated forecasting en

Speaker: Michele Trovero

The abstract: A well-designed forecasting support system, in the form of a commercial package or based on open source, is essential to enable analysts to generate reliable and robust forecasts in a production environment. In recent years, the availability of options for forecasting support systems has been steadily growing and expanding. There are two main drivers of this growth: the evolution of technology common to much of the software industry, and the evolution of algorithms and methodologies specific to forecasting. This talk explores these two elements of change and tries to extrapolate some common trends that will drive innovation and growth in forecasting software for the next few years.

Bio: Michele Trovero is currently leading the Forecasting R&D group at SAS. He has over 17 years of experience in the development of commercial forecasting software. His interests cover a wide variety of topics related to forecasting, time series analysis, and machine learning. He has presented at several international conferences on these topics. He received a PhD in statistics from the University of North Carolina in Chapel Hill and has a degree in economics from Bocconi University in Milan, Italy. When he’s not thinking about forecasting or driving his children to activities, you can find Michele running or riding his bikes.

Speaker: Thomas Willemain

The abstract: Time series scenarios are artificial data series intended as inputs to decision processes. Those processes may involve forecasting, system design, or operator training. We outline criteria for evaluating scenario generators, explain how to create scenarios using nonparametric bootstrapping, and show how to exploit scenarios at the intersection of forecasting and inventory optimization.

Bio: Dr. Tom Willemain is Professor Emeritus of Industrial and Systems Engineering at Rensselaer Polytechnic Institute, Cofounder and Senior Vice President of Smart Software, Inc., and former intelligence officer. He holds degrees from Princeton University and Massachusetts Institute of Technology.

Speaker: Nicos Pavlidis

The abstract: Model combination (averaging) methods have gained substantial attention in the time series forecasting literature due to their success in numerous and diverse applications. In this work we focus on methods capable of accommodating changes over time in both model parameters as well as in the optimal model combination. This problem is particularly relevant as a growing empirical literature provides strong evidence in favour of structural change in numerous application areas. Ignoring the time-varying nature of real-world applications can have detrimental consequences for inference and forecasting. The approach we discuss relies on Dynamic Model Averaging (DMA): an adaptive methodology that is computationally efficient and allows the user to consider a very large pool of candidate models. We discuss modifications of the DMA methodology and in particular the consideration of methods from the machine learning literature for the model averaging step. We apply our proposed approach on time-series from house market prices and compare its performance to state-of-the-art methods for point and density forecast.

Bio: Nicos Pavlidis has studied economics in King's College, Cambridge University. He has an M.Sc. and Ph.D. in mathematics and computer science from the University of Patras, Greece, and has been a research associate at Imperial College London. Since 2010 he has been with the Department of Management Science, Lancaster University. His main research interests are classification and clustering in dynamic and information rich environments, as well as dimensionality reduction.

Speaker: Stephan Kolassa

The abstract: How do we know our forecasts are good, or at least better than some other forecast? There are many forecast accuracy measures in use, some of which come with unexpected side effects. We will discuss accuracy measures for point forecasts, both for central tendencies and for quantiles as necessary for safety stock and capacity planning, and also touch on external industry forecast accuracy benchmarks and tests for statistical significance. If the participants find this interesting, we will organise a follow-up presentation, which will treat interval and density forecast evaluation as well as (as a bonus) the evaluation of predictive classifications. So, please, leave your feedback after the event.

Bio: Stephan Kolassa is a Data Science Expert at SAP Switzerland. His responsibilities include the algorithmic, statistical and forecasting aspects of SAP’s retail platform CARAB, from user research across prototyping to training. He also does some academic research on the side, serving as an Honorary Researcher at the Centre for Marketing Analytics and Forecasting at Lancaster University Management School, and as Associate Editor for Foresight: The International Journal of Applied Forecasting.

Speaker: Stephan Kolassa, SAP

Abstract: One point where we are all confronted with the need for good forecasts is when we go (or nowadays, click) shopping. Retailers need good forecasts to ensure their shelves are well stocked, but not overstocked. (They need forecasts for other uses, as well. More on this in the presentation.) We will go into key challenges in retail demand forecasting, how to address these, what we can learn from the recent M5 competition that used actual retail data, and why UK retailer My Local went out of business.

Bio: Stephan Kolassa is a Data Science Expert at SAP Switzerland. His responsibilities include the algorithmic, statistical and forecasting aspects of SAP’s retail platform CARAB, from user research across prototyping to training. He also does some academic research on the side, serving as an Honorary Researcher at the Centre for Marketing Analytics and Forecasting at Lancaster University Management School, and as Associate Editor for Foresight: The International Journal of Applied Forecasting.

Speaker: Simon Spavound

The abstract: Many businesses rely extensively on their forecasting systems in order to make decisions which concern the future. Given the importance of correct decisions to businesses, a great deal of emphasis is placed on the statistical accuracy of these systems, with little thought in the wider implementation challenges and overall utility within business processes. These issues can create larger barriers to implementation success and crucially user acceptance. With the advent of Machine Learning based forecasting approaches, and business decisions being made off the back of them, trust and the use of these systems by individuals and teams will become ever more important. This talk will discuss these issues, the wider issues of implementation and suggest some solutions to the problems.

Bio: Simon is Data Science Team Lead in Peak. He has PhD in Economics from Lancaster University and is an expert in forecasting in supply chain.

Speaker: Nikos Kourentzes

The abstract: Companies have to produce forecasts for hundreds of up to several thousand products to support their operations. As different functions in an organization plan on different horizons and aggregation levels of the market, from individual products to complete markets, they invariably need different forecasts to support their activities. Traditionally, these forecasts are produced independently, considering each item separately from the rest, without capturing cross-effects between products or sharing information between functions effectively. This can lead to multiple different outlooks for the future, misaligned decision making, and eventually either increased costs or lost opportunities. At the crux of this forecasting challenge is the notion of coherence, where forecasts made at the disaggregated level, for example, for individual stock keeping units, must agree with the forecasts at brand, market, or even more aggregate levels. Although this is intuitively clear, both modelling challenges and organizational frictions can make achieving this difficult. In this talk, we will explore how hierarchical forecasting can help organisations produce coherent forecasts, providing a common view of the future from the disaggregate to the aggregate levels in an organization. Building on this, we will look at how innovations in hierarchical forecasting can help us not only achieve better forecast accuracy, but more efficiently use the available information in an organization, and eventually overcome organizational silos by relying on analytics. We will finish the talk with a practical roadmap to incorporating hierarchical forecasts in organisations, and an outlook of the future.

Bio: Nikolaos Kourentzes is a professor at Skövde University in Sweden, and a CMAF veteran.

Speaker: John Noguera

The abstract: Statistical process control for autocorrelated processes have been addressed using the EWMA (Exponentially Weighted Moving Average) one-step-ahead forecast or simple ARIMA (Auto-Regressive Integrated Moving Average) models. The time series model forecasts the motion in the mean and an Individuals control chart is plotted of the residuals to detect assignable causes. Failure to account for the autocorrelation will produce limits that are too narrow resulting in excessive false alarms, or limits that are too wide resulting in misses. The challenge with this approach is that if there is seasonality or negative autocorrelation in the data, the user needs an advanced level of knowledge in forecasting methods to pick the correct model. In this session, we will review simple exponential smoothing / EWMA and then introduce recent developments in time series forecasting that use automatic model selection to accurately pick the time series model that produces a minimum forecast error.

Bio: John Noguera is Co-founder and Chief Technology Officer of SigmaXL, Inc., a leading provider of user-friendly Excel add-ins for Lean Six Sigma tools, statistical & graphical analysis and Monte Carlo simulation. He leads the development of SigmaXL and DiscoverSim with a passion for ease-of-use, practical & powerful features, and statistical accuracy.  John is a certified Six Sigma master black belt and was an instructor at Motorola University. He has authored conference papers on Statistical Process Control and Six-Sigma Quality and is a contributing author in the Encyclopedia of Statistics in Quality and Reliability (Wiley).

Speaker: Ivan Svetunkov

The abstract: Dynamic model is the model that represents the behaviour of an object over time. The classical examples of dynamic models include Exponential Smoothing, ARIMA, time varying parameters regression and some other models. The three mentioned above are considered to be different classes of models in the academic literature. They have some intersections, but are not necessarily directly connected. Over the last few years, I have been working on theoretical development of state space model that unite these approaches in one unified framework and on implementing this model in an R function. This has resulted in the appearance of the so called “Augmented Dynamic Adaptive Model” – ADAM. In this presentation, I will explain the basics of ADAM and show on several case studies, what problems can be solved with it.

Bio: Ivan is a Lecturer of Marketing Analytics at Lancaster University, UK and a director of Marketing of CMAF. He has PhD in Management Science from Lancaster University. His main area of interest is developing statistical methods for forecasting. He is a creator and maintainer of several forecasting- and analytics-related R packages.

Speakers: Sarah Darin and Erik Subatis

The abstract: Just as the COVID-19 pandemic has disrupted all our lives, it has had a major impact on virtually all businesses. While some businesses have seen demand surge, others have seen it plummet. While some businesses continue to observe a disruptive influence on demand, others have seen demand stabilize to a new normal. Many companies continue to struggle with stockouts and longer lead times.

This webinar will provide pragmatic insight into how to use Forecast Pro’s methods and techniques to create accurate forecasts during the global pandemic. Drawing upon their extensive expertise in forecasting approaches that have been successfully used during this and other business disruptions, Sarah and Erik will provide an overview of how Forecast Pro can be used to account for the impact of Covid-19 in statistical models and how to efficiently integrate judgmental overrides in large scale forecasting projects. Erik and Sarah will profile several business categories, review the methods that are best suited to each group and demonstrate these approaches in Forecast Pro TRAC using real-world examples

Bios:

Sarah Darin is a Vice President at Business Forecast Systems. Sarah received her undergraduate degree in Applied Mathematics from Harvard University and has a M.S. in Statistics from the University of Chicago, where she was a PhD candidate (ABD) and served as a Lecturer.

Erik Subatis is Sales Director at Business Forecast Systems. Erik started at BFS as a product specialist and has been Sales Director for the past 8 years. Erik has a B.A in Writing from Ithaca College and expects to earn a Master’s in Software Engineering from Harvard Extension School next year.

Speaker: Robert Fildes

The abstract: Forecast value added is a phrase now very much part of the organisational understanding needed to improve forecasting practice and to select between alternative forecasting methods. One of its more important applications is to understand the improved accuracy achieved by the common practice in demand planning of adjusting a statistical forecasting method based on information gathered through the sales and operations planning process. Various past studies have analysed the results of this adjustment process. This presentation considers a range of data sources to provide insight into the circumstance where gains have been achieved. We identify the key questions facing any organisation where FVA of expert judgmental adjustment is part of the process. But be warned, the conclusions will not be unequivocal; FVA is a complex area and there remains a lot to be learnt and much to be done if organisations are going to develop effective demand planning processes.

The introduction to this talk was given by Mike Gilliland from SAS.

Bio: Robert Fildes is the Director of Centre for Marketing Analytics and Forecasting, Lancaster University Management School