The described impact has resulted from collaborative research between the Smart Technology Research Centre from Bournemouth University (BU) and Lufthansa Systems (LSY). The efforts focused on improving the quality of forecasts generated within LSY flagship software products, ProfitLine/Yield family (O&D and Rembrandt), used by over 40 airlines worldwide. The resultant forecast fusion model which became a core of the developed software module (ProfitLine/Yield O&D Market Sensitive Forecaster), allowed achieving considerably higher forecasting accuracy. While the exact numbers for individual customers are confidential, it has been estimated that the improvement in forecast accuracy offered by this solution will result in increase of revenue by more than 1%. While the focus of the software is revenue management, the improvement in load factors (capacity utilization) of the airlines which have purchased it, translates to better fuel consumption efficiency and as a result, reduced CO2 emissions per passenger carried.
Details of the impact
Revenue Management Systems are used by airlines, hotel chains, car rental companies, broadcasters and other perishable goods providers worldwide. They were originally introduced by major airlines (e.g., American Airlines and Delta Airlines) in the 1980s, resulting in annual revenue increases of 300-500 million dollars (see Competitive Advantage via Quantitative Methods or Yield Management at American Airlines for further details).
Forecasting is a crucial factor in Revenue Management. In the case of airlines, forecasting of the anticipated booking, cancellation and no-show numbers has a direct impact on general planning of routes and schedules, capacity control for fare classes and overbooking limits, which in turn affects the revenue. The accuracy of predictions is extremely important. It has been estimated by the industry experts that in the airline industry 2-4% of expected additional revenue is generated per 10% of reduced forecast error.
The Combining Predictors project provided theoretical evidence of why and under which conditions multistep, multilevel forecast combination can be a powerful approach in order to build a high quality, adaptive forecast system. The theoretical analysis has been supported by experimental results, which demonstrate an improvement of forecast accuracy by up to 12% for the practical application of demand forecasting in Revenue Management, compared to optimised forecasting system offered by Lufthansa Systems at that time. In the case of mid-size and low-cost airlines, the users of ProfitLine/Yield Rembrandt software, this translates into 2-5% higher revenues, which is equivalent to up to 6 million dollars of additional revenue per year.
The forecast combination and meta-learning methods investigated in the Data Mining and Multi Level Combination for Cancellation Forecasting project resulted in improvement of the booking forecast accuracy by 7.3% with respect to the approach offered by Lufthansa Systems at that time.
The overall effect of applying the advanced multilevel adaptive forecast combination methods and the meta-learning approaches investigated in the above projects was the improvement of overall booking forecast accuracy by 15% in total, as estimated using real airline data provided by Lufthansa Systems. Note, that this value is not a sum of improvements resulting from individual projects due to various compensating effects. According to the airline industry experts the 15% booking forecast accuracy improvement translates into 4-6% of additional revenue for the users of the software, with the exact revenue increase observed by Lufthansa Systems’ customers being confidential.
Lufthansa Systems has also observed a considerable impact of the knowledge gained during realisation of the collaborative projects on the methods that have been implemented in the Market Sensitive Forecaster, a new module of ProfitLine/Yield O&D – Lufthansa Systems’ flagship Revenue Management solution. Initial benchmarks have shown that this system can generate additional earnings of up to 0.25% compared to the traditional O&D Forecaster. The most complex and central parts of the software use forecast fusion approaches and according to Dr Riedel from Lufthansa Systems they would not have been able to implement a market sensitive model covering such a wide range of diverse aspects without that knowledge and without access to BU’s expertise. Lufthansa Systems also takes advantage of the knowledge gained during realization of the mentioned projects by early identification of non-prospective directions concerning development of new features, which allows to use resources more effectively and optimize the development process.
According to the information from AirlineSoftware.net, the company maintaining and providing a comprehensive airline software database, and official marketing materials of Lufthansa Systems, ProfitLine/Yield O&D:
- is currently used by over 40 airlines worldwide;
- allows to reduce forecast errors by up to 35%;
- allows to gain up to 2% additional revenue with O&D approach compared to segment-based approach;
- allows to gain 1% additional revenue with PNR-based no-show forecasting.
The direct business impact of the collaboration with BU related to Market Sensitive Forecaster is already visible as ProfitLine/Yield O&D with this component has been purchased by Lufthansa Passage, Southwest Airlines and other major airline customers. While the exact numbers for individual customers are confidential, it has been estimated that the improvement in forecast accuracy offered by this solution will result in increase of revenue by more than 1%. While the focus of the software is revenue management, the improvement in load factors (capacity utilization) of the airlines which have purchased it, translates to better fuel consumption efficiency and as a result, reduced CO2 emissions per passenger carried.
Underpinning research for the above described impact case is based on three projects that have been co-funded by Lufthansa Systems:
- Combining Predictors: PhD project (2002-06) supervised by Prof. Bogdan Gabrys (BU) and completed by Ms Silvia Riedel. The project focused on theoretical and experimental analysis of different types of forecast diversification and combination techniques and automation of the multilevel forecast generation process. The developed techniques have been applied to the problem of forecasting of seasonal factor predictions in Revenue Management for Airlines [1, 2, 3]. Ms Riedel while carrying out her PhD studies (50% of her time) was in parallel working for LSY in Berlin (50% of the time) allowing her direct implementation of the examined techniques and approaches into Lufthansa Systems Forecasting Kernel. Ms Riedel worked for LSY as a full time forecasting expert before and continued working there after completing her PhD, transferring the knowledge gained during her own and subsequent collaborative projects with BU.
- Adaptive Prediction of Price Sensitive, Low Fare Demand for Airlines: a confidential consultancy project (2004-05) completed by Prof. Bogdan Gabrys in cooperation with Ms Silvia Riedel. The project was a result of the on-going collaboration initiated within the above PhD project and the request from LSY to develop a new price and market sensitive low fare demand forecasting product. The consultancy project developed a critical part of the software component which used the forecast fusion model which became a core of the developed software product – ProfitLine/Yield O&D Market Sensitive Forecaster – later released and sold by LSY.
- Data Mining and Multi Level Combination for Cancellation Forecasting: PhD project (2006-09) supervised by Prof. Bogdan Gabrys and Dr Silvia Riedel (industrial supervisor from LSY), and completed by Ms Christiane Lemke. This project was a continuation of the previous collaborative research with LSY and explored possible improvement of the net booking forecast in the Airline Revenue Management system by modifying one of its components – the cancellation forecasts. The proposed methods have been thoroughly investigated using real airline datasets and compared with the latest state-of-the-art in forecasting research, also using data sets of two recent forecasting competitions, thus being able to provide a link between academic research and industrial practice [4, 5, 6]. In addition to the continued, direct transfer of knowledge to LSY, an international time series forecasting competition winning solution has resulted from this research and was described in .
- Riedel, S. and Gabrys, B., 2007. Combination of Multi Level Forecasts. International Journal of VLSI Signal Processing Systems, ISSN: 0922-5773, vol. 49, no. 2, pp. 265-280.
This paper addresses a difficult real-world forecasting problem with large noise terms in the training data, frequently occurring structural breaks and quickly changing environments. The presented completely automatic, multi-level approach benefits from the advantages of learning on different aggregation levels, reducing the risks of high noise terms on low level predictions and overgeneralization on higher levels. Significant forecast improvements of 12% have been obtained when using the proposed multi-level combination approaches.
- Riedel, S. and Gabrys, B., 2009. Pooling for Combination of Multi Level Forecasts. IEEE Transactions on Knowledge and Data Engineering, 21 (12), pp. 1753-1766. DOI: 10.1109/TKDE.2009.18
The paper presents the theoretical analysis and application of novel multistep, multilevel forecast generation structures carried out in collaborative project with Lufthansa Systems (LS). Significant improvement over industry leading, optimised LS solutions has been achieved.
- Lemke, C., Riedel, S. and Gabrys, B., 2013. Evolving forecast combination structures for airline revenue management. Journal of Revenue and Pricing Management, 12, pp. 221–234. DOI:10.1057/rpm.2012.30.
This paper presents findings of a collaboration project between Bournemouth University and Lufthansa Systems. The main aim is to increase net booking forecast accuracy by modifying one of its components, the cancellation forecast and by using diversification of model parameters. The evolution of forecast combination structures is investigated and shown to be beneficial on an airline data set.
- Lemke, C. and Gabrys, B., 2010. Meta-learning for time series forecasting in the NN GC1 competition. Fuzzy Systems (FUZZ), 2010 IEEE International Conference on, pp. 1 -5. DOI: 10.1109/FUZZY.2010.5584001
NN GC1 competition winning solution.
- Lemke, C. and Gabrys, B., 2010. Meta-learning for time series forecasting and forecast combination. Neurocomputing, 73(10-12), pp. 2006-2016. DOI: 10.1016/j.neucom.2009.09.020
This work investigated meta-learning for time series prediction in order to link problem-specific knowledge to well performing forecasting methods. It was later proven very successful by winning the NN GC1 international forecasting competition held during IJCNN’2010 conference (see above). The manuscript has appeared in the list of most downloaded articles of the Neurocomputing journal in 2010.
- Ruta, D., Gabrys, B. and Lemke, C., 2011. A Generic Multilevel Architecture for Time Series Prediction. IEEE Transactions on Knowledge and Data Engineering, 23(3), pp. 350-359. DOI: 10.1109/TKDE.2010.137.
International forecasting competition winning approach following a class of promoted and developed in our centre multi-level, multi-component prediction frameworks and approaches based on diversified ensembles of predictors.