Michael Feindt is not your average software entrepreneur. The founder of analytics firm Blue Yonder, whose customers include Next and German retail giant Otto, can talk in detail about the challenges businesses face with their data, but he is just as happy talking about subatomic particle physics.
Feindt works as a professor at the Karlsruhe Institute for Technology in Germany, though he is not teaching at the moment. “I’ve just started a leave of absence, so I don’t give lectures any more or mark exams, but I will continue to run research teams for at least the next four years,” he says.
To Feindt, big data is a relative matter. Businesses grappling with hundreds of terabytes each year, might want to look to CERN, the European Organisation for Nuclear Research where Feindt has worked as a researcher. Its Large Hadron Collider (pictured above), which famously attested to the existence of the Higgs-Boson particle in 2013, collects 600TB of data every second during experimental events.
To process and analyse such colossal volumes of data, CERN has become the birthplace of a number of significant innovations relevant to business computing, not least the World Wide Web, developed at its Swiss headquarters by Tim Berners-Lee in 1990, and grid computing.
Despite a computing centre hosting 11,000 servers running 100,000 processors, CERN could not possibly process all the data received in its experiments. In any case, the vast majority of it has no scientific value. Instead, CERN uses hardware “triggers” to screen out data from “uninteresting events”, allowing it to reduce the data it actually records down to about 1GBps – still not trivial volumes for any organisation.
To cope, CERN further reduces the data using software triggers before it is stored. It then uses analytics tools to root out patterns that might indicate novel particles or behaviours from the vast amounts of stored data. During the 1990s, Feindt began to experiment with applying neural network technologies to this analytics process. The result was the Neural Bayesian Estimator for Conditional Probability Densities, developed around 2001 and launched as a product called NeuroBayes for Science.
Expert help in data analytics
Feindt founded Blue Yonder in 2008 and the firm still supports the scientific package, which it offers to universities and research organisations for free.
Since then, Blue Yonder’s self-learning algorithms have been helping organisations find relationships in data and help predict future events in industries including insurance, healthcare, retail banking and manufacturing. It uses internal data, such as sales, product data and promotions, as well as external data, including weather records, school holidays and high profile events.
"Many users sit in front of a screen and type stuff in – that is the way most software works. That’s what we don’t want"
Michael Feindt, Blue Yonder
For example, German retailer SportScheck managed to improve accuracy of sales forecasting by between 20% and 40%, helping it to avoid stock outages or carrying too much stock, according to Blue Yonder.
While businesses generally store a great deal of data about their performance, few actually exploit it to the full potential, says Feindt.
“A lot of businesses have ‘write-only databases’, where people put stuff in but no one is reads it out, and no one is capable of making sense out of the data. There is an IT department, but not people who can analyse data to make statistical correlation or understand whether there are causal effects in it. This needs a scientific education, and it is simply not there in business,” he adds.
The supplier offers its algorithms as part of a cloud service. Some are pre-packaged to solve particular business problems, such as demand forecasting, while others are more open ended. The company employs around 80 PhD-level data scientists experienced in using NeuroBayes algorithms who can help customers with data.
“We have software for specific problems, like automated replenishment or customer selection. But you need the expertise to make sure the analysis is applied correctly. As we offer a cloud service the client does not have to have software or hardware or expertise.”
Although businesses might use standard business intelligence (BI) tools, these are not sufficient to understand more complex relationships in the data, says Feindt. He wants to move away from a world where users need to move data in and out of BI and analytics tools to make predictions, towards a model where data is fed directly into automated decision-making systems and the output fed into enterprise software.
“Many users sit in front of a screen and type stuff in – that is the way most software works. That’s what we don’t want. The decision should be done automatically and, for example, the orders should be written back into an SAP system. Existing tools are simply not able to do all of this.”
For this reason, Blue Yonder tries to sell into board-level decision-makers, to argue that this is a new approach, different to existing analytics tools. “We work almost exclusively with C-level. We say to them, ‘Either you make a small improvement, which costs a lot of money, or you get the real thing’. The real thing is so different, so much deeper and more scientific than [standard business software].”
Feindt says that to create this kind of change in an organisation’s decision-making process could would take years if it were started from the IT department, trying to get approval through a hierarchical structure.
“This must be business-driven. IT sometimes likes it, sometimes not, but the largest resistance is not from IT – it is usually from people who have power in the business and who are afraid of losing power if decision-making is automated. But some business managers will see they can have it 20% better and will never be able to do it any other way,” he says.
Seeking causal effects in data
Exactly how Blue Yonder can improve performance depends on the problem in hand, says Feindt. To simply predict the future, by forecasting how much of item X a retail business will sell the next day, building up patterns of correlation between data can be sufficient.
Read more about predictive analytics
- Predictive analytics: The branch of data mining concerned with the prediction of future probabilities and trends.
- Using predictive analytics tools lets organisations look ahead in an effort to optimise business strategies.
- Food retailer EAT uses predictive analytics to cut food waste by predicting future demand better.
“You have to assume what was right in the past 10 years will not be completely different tomorrow. We have learned that the variables that predict future events are generally stable and we find this is true through changes in laws, changes of tax rates and the financial crisis. These relationships are stable – human behaviour does not change from one day to another,” he says.
The problem becomes more complex when an organisation asks what effect changing policy will have on the future. This is the case with pricing, where a change will actually affect consumer behaviour and the policies of competing firms. Then correlation is no longer sufficient; the business needs to model the system to understand what the outcome is likely to be.
“Usually the big data community ignores this. We have developed algorithms which help you can find causal effects in data, not just correlation. So, if you change your policy in the future, you can model what the effect will be. It is quite new and it is extremely sensitive. It is way more difficult than finding correlation, but you can do it,” says Feindt.
If he is right, predictive analytics promises a very different kind of business, where companies can see the effect of their decisions before they execute them. But although the science may be there, the question is whether business is ready.