As online brokerages suffer a shortage of human
resources, kabu.com Securities "hires" AI

Trade surveillance is similar to craftsmanship in that past heuristics are all-important. Suspicious behavior is detected by looking at how many orders are placed at which timing, or from patterns in a series of trades. kabu.com has a limited number of experts who can take charge of trade surveillance. Nevertheless, with the number of trades rising, rogue methods are also becoming more sophisticated.

According to Mr. Saito: "What we want are not novices, but highly experienced people who are effective from day one. Recruiting such people now, when people are hard to find anyways, is hard. Besides, it is basically impossible to search through all of the several million trades that occur daily no matter how large our staff. All things considered, we had no choice but to automate." In an environment of fierce competition for human resources in the brokerage industry, where all the talented people are already employed, kabu.com found a way out of this situation via AI.

As a way to mitigate the shortage of human resources, kabu.com considered introducing Hitachi's AI product, known as "Hitachi AI Technology/H" (hereafter “AT/H”). Says Mr. Saito: "We tried analysis AI, image recognition AI, language processing AI, all kinds of AI from everyone from venture startups to major companies." His reason for selecting Hitachi's AI was its strengths in analysis and figures, enabled by Hitachi's proprietary Leap Learning technology, making it appropriate for the financial operations of a brokerage. kabu.com then began applying it to two of its operations: stock-lending and trade surveillance.

Stock-lending is a transaction in which stock owned by an investor is temporarily loaned to a brokerage firm for which a commission is received. Stock borrowed from an investor by a brokerage is then lent to other institutional investors or supplied as stock for short-selling in order to create revenue. However, even as demand continues to grow each year, kabu.com has a mere three traders negotiating about 500 lending transactions a day. With trade surveillance being done in a similar situation, it became desirable to increase the number of talented traders.

The first operation to which AT/H was applied was the determination of stock lending rates—a particular bottleneck. A system was configured that analyzes over 1,000 types of numerical data and automatically calculates the optimal lending rate. This made it possible to greatly reduce the response time to institutional investors, resulting in a 30 to 35% rise in sales from the stock-lending operation. As Mr. Saito comments: "Probably no one is aware that we are handling this with AI. This success just reinforces my own recognition of AI as a technology for making money."

The next operation targeted for AT/H application was trade surveillance. Adopted by the Tokyo Stock Exchange, AT/H is in use in examination during the initial phase of trade surveillance. This system frees the trade surveillance officer to focus only on detailed full-scale surveillance.

Proof of Concept (PoC) for AT/H application to trade surveillance was carried out on "layering"—a technique that attracts third parties to trades by initiating a large volume of trades without any intention of completing them. The verification process went forward as Ikuo Kurosawa, chief of kabu.com's Trading Screening Section, and Hitachi engineers went through numerous rounds of direct discussions.

The main feature of AT/H is a Hitachi proprietary technology known as "Leap Learning." "Deep learning," a familiar AI technology, is characterized by calculations that, like the interior of a black box, are unknowable. In contrast, a feature of Leap Learning is that it expresses via a model what kinds of reasons are behind certain results. (Figure 1)

Specific trade conditions used in judging suspicious trade activity, such as [trader age of between 20 and 30] plus [number of trades per day exceeding 100], are shown in the model. For this reason, even if AI was not able to detect unfair trading during the learning process, this serves as a clue in considering the kinds of conditions setups that need redoing. In addition, having AI learn new methods of sales transactions allows the model to be tuned, making it possible to approach areas handled by people with intuition and years of experience.

Hitachi learned from direct interviews with Mr. Kurosawa the kinds of patterns that had a high probability of falling under “unfair trading.” Based on this, Hitachi then formulated a hypothesis and generated an unfair trading determination model using AT/H. Trade data for the verification process was input into this model, which then performed scoring for the probability of unfair trading. After the results were checked by Mr. Kurosawa, the PDCA cycle was executed in order to redo the conditions setup. As a result, scoring by AT/H made it possible to exclude 73% of the trades that did not need surveillance such that the trade surveillance officer had only to look into the remaining trades. Next, PoC was executed for cross-trading, in which the same officer initiates sell and buy orders at the same time and price. This resulted in being able to exclude over 90% of all trades. (Figure 2)

After six months of PoC work, the system was implemented in an exceptionally fast 2.5 months, with full-scale operation launching in August 2018. As Mr. Kurosawa notes: "In the face of tough demands, some of our discussions were more like arguments." The application of AI in such a short period can only be attributed to how Hitachi engineers and kabu.com's workplace leaders worked closely together against a backdrop of sincere discussion.

Mr. Kurosawa describes the benefits of implementation as follows: "The workload on everybody was drastically reduced, and because it felt like an injection of new staff members into the team, the overall feeling of the team got radically better." Emphasizing the unique merits of Hitachi AI featuring model indicating, he adds that "AT/H lets us visualize know-how by indicating in a model customer behavior that is highly correlated with layering. This is the same thing as receiving training from an experienced trade surveillance officer."

"The presence of a trainer who trains AI is of upmost importance. Treating AI with the same care as a valued apprentice will produce huge leaps forward in implementation benefits," says Mr. Saito. The AI will gain know-how equivalent to that of experienced employees and will convey that to the next generation. It is as if the company simultaneously acquired both a talented employee and a teacher.