Goods4Cast: smart demand forecasting

TAdviser: What is the Goods4Cast platform? Alexander Yefimov: The Goods4Cast platform is own development of the company and first of all is directed to the solution of a problem of high-quality planning of trade inventories for the trading companies.

The mathematical basis of the platform is formed by the last achievements of the Russian science in the field of forecasting. Development and improvement of prediction algorithms takes place under the leadership of Candidate of Physics and Mathematics of the associate director on researches of Konstantin Vyacheslavovich Vorontsov and the doctor of physical and mathematical sciences of the director of science Konstantin Vladimirovich Rudakov.

The software solution on the Goods4Cast platform contains a number of the components integrated around the predictive server. The final solution is completed individually under the client and depending on requirements can include components for calculation of the recommended order, operating control of goods availability on shelves, accounting of bonuses to selling assistants, optimization of volumes of safety stocks, including taking into account the events of default of orders suppliers, efficiency evaluations of promotion actions and other impacts on demand.

TAdviser: In what industries do you see applications of your solution? Problems of demand forecasting and planning of stocks are relevant not only for retail, but even for distribution and production. Alexander Yefimov: Historically the Goods4Cast platform was developed for retail chain stores which in connection with a large number of goods have a problem of the automated resuming of a remaining balance in shops. Nevertheless, solutions based on the Goods4Cast platform have broader application, and are used for solving of tasks in distribution and production.

One of examples is the problem of sales planning which solution requires high-quality demand forecasting and accounting of the objects set for the company. For formation of the reasonable sales plan first of all it is necessary to make a start from the requirements of the market predicted on the basis of sales. Primary plan received thus, as a rule, undergoes several iterations of approvals and adjustments taking into account external restrictions and according to the purposes of the company. As a result the final sales plan forms. For support of process of approval and recalculation of the plan approach of what-if of the analysis by means of which it is possible to answer questions is used: as sales on a commodity group at introduction of new goods will change as introduction of a promotion action or the change in price will affect sales.

The answer to these questions can be given only by means of creation of the forecast taking into account the various factors and external restrictions imposed by feature of a task and business processes of the company.

TAdviser: What companies already use your development and is how successful? Alexander Yefimov: The system of automatic resuming of a remaining balance in shops was developed and for the first time implemented in Perekrestok trading house. The project success predetermined development of the Goods4Cast platform and further implementations of this technology. Today Goods4Cast is set in retail chain stores Svyaznoy and Lama (grocery chain in Tomsk), also our solution is used in the Baltika PC for formation of sales plans and production.

For efficiency evaluation from system implementation of resuming of a remaining balance the special technique is used. According to this technique the losses from deficit frozen in a remaining balance of goods of means, loss from write-offs on a goods expiration date are subject to control (for food products). On total losses which consist of the specified indicators system operation of Goods4Cast and the client's algorithm used earlier is compared. Reduction of losses after system implementation of automatic resuming of a remaining balance of Goods4Cast averages 1.5% of turnover.

TAdviser: Today in many retail chain stores ERP systems in which there is a forecasting module are already used. What it is possible to justify need of implementation of one more, additional system of demand forecasting with? Alexander Yefimov: The basic purpose of the ERP system is an automation of accounting and management in the company. The possibility of forecasting is only the ERP systems add-on module therefore often in such modules rather simple prediction algorithms based on the speed of sales are used. However it is required not only to apply more competent prediction algorithms to creation of the qualitative and reasonable forecast, but also to be able to consider a big variety of factors which influence demand. When developing the Goods4Cast platform we focused on forecasting solving of tasks.

It should be noted that the companies which implemented the ERP system have a certain advantage: the system of forecasting at them will be based on already smoothly running processes of data collection. At solving of tasks of data mining, including forecasting, the lion share of success depends on quality of collecting and accounting of data. Therefore it is possible to tell that implementation of the ERP system is an obligatory step before implementation of any BI solution.

TAdviser: You are leaders of the Russian market of the systems of demand forecasting. Whether the western competitors disturb you? Alexander Yefimov: What now occurs in the market, demonstrates that this market is rather nedoosvoyen, than is oversaturated by a large number of the competing solutions. It is possible to tell that we work for market development of the systems of forecasting hand in hand with other developers, implementing a thought that such systems are necessary and useful in consciousness of potential customers. On the other hand, in comparison with the competing systems we have rather serious advantages – our system was developed under the Russian market while the western solutions are created taking into account specifics of the western markets which are at other maturity. Respectively, under our conditions it is easier for us to give correct and adequate forecasts. Also it is necessary to take into account also price factors, and proximity of office of development to the client.

Moreover, we had an opportunity to compare our approach to forecasting to the western vendors. In 2004 Perekrestok company and in 2006 Svyaznoy company held open competitions in comparison of forecast accuracy of the systems of forecasting from the leading world IT companies on real data on sales. In both tenders the solution on the Goods4Cast platform won a victory.

TAdviser: What algorithms formed the basis of your system of forecasting? Alexander Yefimov: Our company grew from school of sciences of the academician of RAS Yury Ivanovich Zhuravlev and is formed based on Computer center of the Russian Academy of Sciences (Data center RAS) with which we as the company densely cooperate already nearly ten years. Such proximity to the academic science allows us to develop and use the most advanced developments and achievements of scientists.

Many ask us about what algorithms it is better to use for these or those goods, a time series, group of goods. We apply approach which allows a system to select automatically the best prediction algorithm or to build superposition of algorithms for the purpose of minimization of an aggregate error on the predicted interval. Such superposition allows to take from everyone even if elementary, an algorithm and the quality of the forecast is essential to raise its best sides. In this approach we use both classical and widespread prediction algorithms, and algorithms of own developments.

File:Forecsys-d1.jpg

Illustration of approach of superposition of prediction algorithms.

TAdviser: What do you put in the concept "smart forecasting"?

Alexander Yefimov: Obviously, the ideal forecast is that forecast which matches the fact. However ideally exact forecasts do not exist therefore it is required to define to what losses lead errors of forecasts. The error in this or that direction concerning the fact is inadequate in terms of financial losses. If the actual demand appeared more, than was predicted, and there was out of stock, then losses of the company are the missed profit equal to the value of the margin increased by quantity of not sold goods. If the actual demand appeared less, than was predicted, and there was over stock, then losses of the company consist of the means frozen in excess of goods and cost of the goods which are written off on expiry date (in case of food products). In addition, irrespective of a situation of over stock or out of stock, there are costs connected with storage of goods.

The Goods4Cast system was purposefully developed taking into account financial features of structure of errors and is expressed as target asymmetrical loss function. Such any target function and a method of its optimization is know-how of our company that is not provided neither in classical prediction algorithms, nor in products of the western vendors.

File:Forecsys-d2.jpg

Illustration of asymmetrical target function.

The additional benefit from use of such target function consists in a possibility of effective accounting of policy of management of trade inventories of the client, for example, up to each goods there is an opportunity to set the allowed level of deficit, or from a ratio of indicators of structure of losses at the current price and a margin to make decisions on need of presence of goods at an assortment matrix.

TAdviser: If the system of forecasting is implemented in large retail network, for certain the project will involve big equipment costs which will execute calculations? Alexander Yefimov: This delusion. Our algorithms are very effective, in particular, on rather simple servers (worth up to $10,000). For example, on the server from 2 x Quad-Core the processor (two four-nuclear processors) and RAM 16 GB in one hour make calculation of 250,000 forecasts. Such system performance quite is enough for processing in four hours of 50 shops from 20,000 positions of the active range, at the same time a system uses a five-year sales history.

TAdviser: Any implementation should be justified in terms of efficiency. What can you tell about payback of the Goods4Cast system?

Alexander Yefimov: Our technique of assessment of economic effect of system implementation of forecasting of Goods4Cast is rather transparent. On each of stages of installation of a system calculation expected (is made during implementation and test operation) and the actual system effectiveness (during trial and commercial operation). Total costs which consist of the indicators of deficit of goods frozen in balance, write-offs on expiry date are controlled. Reduction of total costs after system implementation is a performance indicator. In retail chain stores with turnover of $100 million the system of resuming of a remaining balance Goods4Cast pays off in two months after its introduction to commercial operation on all network.

TAdviser: What you see your system of forecasting in the future?

Alexander Yefimov: The software product of a class of BI solutions cannot be complete – it should develop constantly taking into account new requirements. Of course, we have far-reaching plans connected with improvement of quality of forecasting both due to expansion of library of algorithms and at the expense of new accounting methods of additional factors. Also we constantly work on system operation fall forward.

In addition to improvement of the existing product we work on estimating joint influence on demand of mechanisms of pricing and interchangeability of goods. The price, interchangeability, demand – this three lies not so much in the field of forecasting of sales how many treats management of profitability and management of image of network. Changing price level in relation to competitors, it is possible to influence positioning of all retail chain stores, and it, in turn, can lead to change of assortment policy. Whether there is a sense to be guided by the range only in a high price category or to select the range focused on different segments of the population? It is difficult to receive answers to these questions. Here anybody has no ready, finished solution, and the new offer in this area can quite become break.