Overcoming information asymmetries using artificial intelligence
In business life, companies are repeatedly faced with the challenge of assessing the characteristics and behaviour of market partners. In most cases, market partners will not disclose all their information, which in most cases forces the company either to make a more or less blanket assumption about the market partner or to ignore its characteristics and expected behaviour. In case of doubt, both lead to inefficient decisions. The use of artificial intelligence is suitable to reduce the problems resulting from these information asymmetries.
Every company with information disadvantages compared to a market partner is anxious to prevent an exploitation of the information advantages. This effect of gaining an advantage from a knowledge advantage is also known as moral hazard. It can be found in the quantification of default probabilities in the credit business as well as in the mail order business, where the retailer is unclear as to whether a customer actually wants to purchase and retain an ordered product. The “Principal Agent Theory” therefore generally assumes opportunistic action by the “agent” (contractor) between market partners if he has a knowledge advantage over the “principal” (client).
In the past, attempts were usually made to reduce the negative consequences of information asymmetries by skilfully drafting contracts or making assumptions that were as accurate as possible. Statistical means based on past experience are usually used for this purpose. However, these are not always reliable, especially when market conditions change. With today’s means, such as the use of neural networks (deep learning, often also called “AI”), much better assumptions and forecasts can be derived.
If one looks at large Internet companies such as Amazon (“customers who bought this product also bought…”) or Google (customized individual advertising), it seems possible to reduce information asymmetries.
Deep learning allows patterns to be identified that can contribute to the elimination of disadvantages caused by information asymmetries.
A trained neural network is able to recognize connections where the human brain reaches its limits. Neural networks, as they are used today, have an input layer, a plurality of “hidden” layers and an output layer. The network contains a large number of neurons that process the input data and weight it according to the desired output. The network learns independently how to weight the individual factors in order to achieve the desired result. The training and the control of the network are carried out on the basis of real data, which are, however, used separately. This makes it possible to control how valid the output of the network is.
Derive use cases and increase potentials
Companies facing the challenges of asymmetric information distribution should identify use cases in a structured procedure and ask themselves the following questions:
- In which areas are there regular disadvantages due to missing or incorrect information?
- Are there any data that allow conclusions to be drawn about the missing information and how can these be collected/related?
- How can the newly acquired information be used sensibly?
- What would be the benefit of closing the information gap?
The quantity and quality of the data to be used to eliminate information asymmetries have a significant impact on the expected costs and benefits of using machine learning and deep learning. The input data must be consistent in itself. In addition, it is advantageous if the main influencing variables on the output are also known.
Optimize overall bank management with Deep Learning
The questions posed are equally relevant for companies inside and outside the financial industry. With the data available to them, however, banks usually have an exceptional starting situation. In addition to the master data of individual borrowers, the bank also has information that could be relevant for overcoming information asymmetry. The task of an AI solution is to recognize the relevance as well as the correlations of individual variables for the challenge that arises.
In addition to a person’s age, gender, marital status and occupation, the bank also knows the individual input and output situation. From salary and monthly rent payments to other obligations, the bank can assess a person’s financial situation fairly well and identify medium to long-term developments. Expressed in figures: A medium-sized bank usually has 400 slightly different information per account per key date, usually at the end of the month. That’s a good 4,800 records per year per account. Institutes today use only a fraction of these.
Neural networks can help to evaluate this data: For the construction, a bank defines a certain number of input variables and feeds them to the model in the input layer. Each of these variables converges in one or more nodes (hidden layer), which, depending on the weighting of the individual variables, leads to a specific result (output layer). A realistic result requires training and a validation record.
This information can be used to cover a wide range of use cases from the various disciplines of overall bank management: For sales controlling, it is possible to determine the affinity of a customer with regard to the bank products required or the sales channel, thus making sales more efficient and improving the customer experience. In liquidity and interest rate risk management, the exercise probabilities of optional components such as special redemption rights can be better estimated and cash flows more accurately modelled. And in credit risk management it is possible to determine better statements with regard to default probabilities. The possibilities are almost unlimited.
In addition to master data, current and credit account transactions over as long a period as possible are suitable for most use cases. It should be noted that the cross-section of the portfolio should be shown. Approximately 70 percent of the data set is used to train the model, while the remaining 30 percent is used for subsequent validation of the model. It is important that the relationship between the customer groups is the same in both data sets.
Procedure for training a neural network
The interesting thing about Deep Learning is that not all relevant variables or their influence on a result necessarily have to be known and the model can be extended almost arbitrarily. For training purposes, it is advantageous to know the influencing factors that are highly likely to have a significant influence on the result. The weighting of the individual variables among each other is carried out independently by the model on the basis of the training data set.
As output, a bank usually receives two types of results: a statement as to which result can be expected and which factors influence this result. With this information, banks can not only better quantify unknown variables, but also implement an early warning system that shows key risk drivers and their changes. The use of Deep Learning is thus the optimal support for active management of challenges that banks have had to face until now.