3 questions to Dr. Stefan Ebener

How hard it’s for people when AI enters their working life? What are the main reasons for reservations or even rejection?

Ebener: It is uncertainty, a difficult situation for our brain that can influence our rational way of thinking. This uncertainty and resistance to change are largely due to the fear of losing security or control, competence and recognition, and relationships. The main reasons for reservations or even refusal can therefore be best associated with it.

However, the vast majority of us already work with applications that are equipped with artificial intelligence. Most of us simply do not know that artificial intelligence supports them in their actions. So my gmail account is organized by an artificial intelligence and offers me for each e-mail three text modules as small buttons to send a reply in a few seconds. If I write a more detailed mail, the system completes the sentences I started and even understands the context. Never in my professional life have I answered e-mails so quickly - let alone the writing.

The same applies to the use of AI in many other areas of work. The perceived loss of security and control, the ignoring of one's own gut feeling, lead to reservations and the perception of risks, where actually benefits and opportunities clearly outweigh. Like so many technological developments over the past 75 years - the TV, the Internet, the smartphone - the desire for openness to new technologies is the key to the future.

What is the biggest "stumbling block" when forecasting with AI?

Ebener: AI-based forecasting systems have their own challenges and should not be used without a strong understanding of the market, especially in the financial sector. To get into the topic, you should a.g. to deal with the following "stumbling blocks":

1. AI is more than neural networks: Data scientists tend to solve problems directly with neural networks. Neural networks are powerful, hip and have the reputation of being able to solve everything. All too often it is neglected that new problems should first be tackled with simple, comprehensible models. This can also be the basis for the performance of sophisticated models.

2. Be careful when using AI on "low frequency data": The frequently used KI models in controlling need high-frequency data and information ("high frequency data"). Otherwise there is a danger that the neural networks will memorize the data instead of understanding the relationship of the data. Established Arima models use data in too low frequency to use for evaluation with AI. Therefore: Stay away from "low frequency data" in combination with neural networks!

3. Technology alone is not enough - background knowledge is indispensable: there are now countless models (and companies) that promise perfect forecasting systems. These systems are based partly on sophisticated models, which can be very dangerous without the corresponding background knowledge. Thus, some models are very much inclined to make connections that do not exist. Without appropriate background knowledge, this can hardly be recognized. In the right hands and with the right knowledge, however, these can be very powerful tools in identifying important causalities.

4. Always critically question backtesting in recent years: Since 2008, we have been living in a fairly stable financial position with steadily rising prices. This can be described as a special market situation. In the development and testing of new AI models, you should pay particular attention to the balance of the data, for example, for the so-called backtest. For the last ten years in the financial sector have been extraordinary and could lead artificial intelligence to make bad recommendations.

5. No help with shocks: 2008 showed that certain models were able to maneuver relatively well through the crisis - especially in terms of the alpha, the excess return to the market. One reason for this was the continuously increasing volatility of the summer 2007. "Value at risk" models were one of them. Contrary to popular belief, however, even the most sophisticated AI models struggle with the prediction of shocks. Frequently, AI models respond to even more changes than would be appropriate in the situation.

What do you advise controllers in particular?

Ebener: Do not be afraid of AI systems! Use them as an additional medium. Finances are always hard to predict, even with AI.

I'd like everyone to keep in mind that there are special challenges in financial forecasts. The problem is completely different than, for example, in image data analysis. Here, the data between training and test are relatively similar (cat pictures include cats). In the financial sector, future events may have completely different distributions of data. Among other things, the "walk forward" method is used here, which in turn makes assumptions about the future distribution and thus accepts losses in accuracy: 56 percent accuracy is therefore the maximum for financial forecasts, as opposed to 99 percent in image analysis.

Forecasting also has a special challenge: the future is uncertain and yet a trading decision has to be made today. The information available is often minimal and the data distribution changes at the same time. „Reinforcement learning”, a form of machine learning, addresses this multi-dimensional problem and could help to provide better models in the future.

About Dr. Stefan Ebener

Dr. Stefan Ebener is the manager of Customer Engineering for Google Cloud and leads an international machine learning and AI expert team. His passion is the data-driven future technologies and the development of technological competencies in companies and society to which he regularly appears as a keynote speaker.

In addition, he is a freelance lecturer in business informatics and deals with ML, AI and big data with the topic "Opinion Leader Identification & Management". His research interests include metadata and text mining, machine learning, as well as competitive and tender analysis.

As a trained data scientist, he has practical experience in building data pipelines and model development. Ebener is also involved in the research area "Business Analytics" at IFID, the Institute for IT Management & Digitization at the University of Economics & Management.