An Interview with Yoshi Yokokawa, the CEO of Alpaca, by Longine Fintech Team
We asked AlpacaDB’s CEO, Yoshi Yokokawa, about Alpaca’s service and underlying technology, as well as their business development prospects.
Three Key Messages to Readers
Alpaca, as well as its predecessor, Ikkyo Technology, was founded by three university classmates.
Capitalico is a tool that enables users to automate their own trading techniques and ideas without programming. Once automated, users can also back-test and quickly observe the results.
While it utilizes image processing and deep learning, the data necessary for trading is limited. Alpaca works through trial and error, asking the users themselves what they can do.
Leading to the Launch of Alpaca DB
Longine FinTech Interviewer (hereafter “Longine”): Congratulations on the FIBC2016 Award. Before we ask about the service that received the award, please tell us how Alpaca got started.
Alpaca’s CEO, Yoshi Yokokawa (hereafter “Yokokawa”): The story dates back three years. Alpaca’s current founding members launched Ikkyo Technology in February 2013. “Ikkyo” is the name of an Aikido technique, which literally means “one teaching,” referring to a basic form of Aikido. Yuki Hayashi (Alpaca’s Chief Engineering Officer), who co-founded the company, is an Aikido master. We named the company Ikkyo, determined to create a technology company.
Hayashi has a background in developing large-scale distributed network systems and the core modules for 3-D graphic-related products. He originally specialized in these at Nintendo and Nokia during a time when mobile games were becoming crazily popular. When we started Ikkyo, albeit in a different field, we utilized his expertise in this distributed processing and ultimate acceleration to do things like creating a distributed in-memory database that spits out sales reports from a major pharmaceutical company at a speed several ten times faster than before. But we struggled because the business does not grow fast when it is mainly based on contracted projects.
Longine: What happened to Ikkyo then?
Yokokawa: We launched Ikkyo in February 2013 and restructured it in February 2015. Once we launched Alpaca DB in Delaware, in the United States, we transferred the shareholders from Ikkyo. Today, Ikkyo is a Japanese subsidiary of Alpaca.
Longine: Please tell us about the engineers who are the founding members.
Yokokawa: It was founded by the three of us, but the two others who were engineers unlike me, were best friends from the university. Hayashi, whom I mentioned about earlier, was already a programmer when he was in elementary school. He was like a hacker. He joined Nintendo after graduating from university, but had already been a technology officer for an industry-academia collaboration between Internet companies and the Keio SFC (the Keio University Shonan Fujisawa Campus). He used to write visual engine programs in assembly languages for domestic telecom operators too. When he was in junior high school, he used to create small programs with images, video, and sound in a demoscene, called “megademo.”
Hitoshi Harada (Alpaca’s Chief Technology Officer) specializes in database technology. He began working as an intern at a system development company while in college, and had developed search engines for airline companies and travel websites. Around that time, he was also involved in the PostgreSQL community as a hobby, and was quite famous. The EMC’s technology division head-hunted him because he became known throughout the world for that, and he joined EMC’s Silicon Valley office. As EMC had acquired a parallel and distributed database company, called Greenplum, around that time, he worked as an architect of Greenplum’s team. Now one of the co-founders of Greenplum works for our company.
Changes after the Establishment of Alpaca
Longine: What kind of changes did you experience as a company when the management foundation transitioned from Ikkyo to Alpaca?
Yokokawa: I used to sell technology for contract work, but I began to adopt an approach of creating a product to meet demand. We created a machine learning product for image recognition, kind of like creating the greatest common factor, thinking “perhaps it could become a product if something like this existed.” As deep learning also came along simultaneously, we tried to create a web-based image recognition service that would utilize that. That’s the service called “Labellio.”
Longine: So the prototype led to the product.
Yokokawa: It was released in June 2015. However, as I struggled to sell it, it became difficult for me to remain motivated, so I talked to the founding members, Hayashi and Harada. As we were stuck, the three of us decided to get together to discuss.
I had worked for Lehman Brothers and Nomura Securities after I graduated from university. After I quit Nomura Securities, I day-traded in the Forex market for about three years. As the very day-trading environment left a user-unfriendly impression on me, we ultimately decided to provide a solution to improve such an environment for day-trading.
Longine: What led you to day-trading initially?
Yokokawa: If you research thoroughly, financial products exist that allow you to obtain a considerable alpha, but these are often products available only for institutional investors. Even if I want to buy it personally, it’s often not available. I began day-trading because I thought the forex market offers investment opportunities in a realm that is not dependent on the market’s beta. It’s also highly liquid and easily accessible to individuals. In other words, day-trading was my solution for asset management.
Longine: How did day-trading lead to the idea for Capitalico, which won the FIBC2016 Award this year?
Yokokawa: The key to successful day-trading is really a series of patient practices to develop professional skills. Trading itself also includes the element of repetitive practice. There is a significant psychological stress in thinking, “I might want to exit when the money trades this way,” or not being able to cut a loss, thinking, “It will bounce back in a while.”
As you frantically endure, while crying because it doesn’t easily go well, something of a model algorithm emerges that makes you think, “Oh, doing this might work.” However, when you actually try trading using your own model, you now feel stressed about following that rule, remembering that it is painful.
You think, “Why can’t this be done by a machine?” and try to automate your own model. You try to create an Expert Advisor (EA), called MQL (Meta Quotes Language), but the program won’t work at all as you envisioned. I’m not capable of fine-tuning or creating machine learning either, as I’m not a programmer. I ended up day-trading by eyeing the market for more than ten hours a day, seven days a week. This is why I want to eliminate such pain.
I told them how I wanted to create this solution at our company, and we discussed it among the three of us, as the founding members. We then decided to make this idea happen and, after checking if the image processing technology and the base technology for deep learning can be applied to trading, and changing the course of direction multiple times, we launched the service as Capitalico.
Aim of Capitalico
Longine: Please tell us again about the details of the Capitalico service.
Yokokawa: It is a trading tool, and specifically, it’s a tool that allows you to take the trading techniques and ideas you are using and automate them without programming. Once automated, you can also back-test and easily observe the results. Additionally, we are hoping to turn it into a marketplace, where you can sell what you have automated to a third party.
Longine: Does this mean it will become a trading tool that is quite open?
Yokokawa: Once that happens, the users of our service will include not only those who create trading tools, but also those who want to use the tool to invest, so we had planned to include those in one app. We changed our direction toward mobile because we felt it was necessary to provide a mobile app, rather than a web browser version, to do so. The latest development is that we released an iOS app on March 11, 2016.
Longine: What kind of reactions are you receiving?
Yokokawa: The actual reaction, in fact, turned out to be different from our expectations, and we began seeing some issues. That is, a situation in which it falls short for those who want to create an elaborate trading tool, but overkill for those who are interested, even though they don’t understand technical analysis. We decided to focus on the segment of ultra-professional people who really get into this, from now on. As a result, we envision a service to be used on a web browser, or on a desktop, rather than on mobile devices.
Technical Background of Capitalico
Longine: I know Capitalico uses image processing technology, but would you tell us how it actually uses the technology?
Yokokawa: What we realized when we were using image recognition in the past is that there are two points: whether to create a general-purpose product, or to automate to the judgement of each individual. Google takes the approach of creating a robust AI for general purposes, but what we realized when we were working on contracted projects on image recognition was that every issue is different.
For example, you might say, “Create a system that can recognize certain Japanese children,” but we cannot really find one-size-fits-all types of sample data on the faces of Japanese children on the Internet. As a result, it becomes difficult for Google to learn what it is. If you try to have Google determine the sex of chicks, for example, it’s not possible unless a skilled Japanese chick sexer teaches Google.
Longine: How do you draw a line between what AI is good at and not good at?
Yokokawa: It may be similar to the difference between the eye of the general public and that of experts. With trading, we focus on the eye of experts, or rather, the way the experts think. On the other hand, while everyone is making much of AI, the majority of attention is directed, in fact, to the approach for creating generic, robust AI. Although this is an approach to feed as much data as possible and hope to obtain guidance without inputting any human judgement, it seems that we are heading in the other direction.
Longine: How about deep learning?
Yokokawa: The deep learning category includes various types. Image recognition uses a convolutional neural network, and what we are primarily using now is based on a recurrent neural network. Additionally, we look at all the latest methods proposed by academia that could work with a little arrangement, and test them systematically as they come out. We applied reinforcement learning, such as the one used in AlphaGo, to optimize strategy, and attempted to generate pseudo-chart data using a “generation model” network to resolve the shortage of training data, for example. Incidentally, Hayashi had been working as an engineer in the gaming industry from early on, and since one engineer typically handled the graphics, sound, gaming system, and AI by himself in the gaming industry at the time, he has been familiar with the neural network-related technology for a long time, albeit not as deeply as today.
Longine: How do you combine image processing with deep learning?
Yokokawa: How to establish a network with deep learning, how to tune up a network, or how to distribute it to avoid over-learning ? these kinds of network knowledge are being continuously developed. Meanwhile, deciding what kind of data to input, such as whether to input images or financial data, for example, makes a huge difference. Unlike images that provide general knowledge, such as “This is a PC” and “This is a cup,” financial data involves teaching user-specific judgements. In such a case, you must create a different, individual-based network, one by one. Additionally, since the type of input data package would vary by individual, you would also have to develop a network that can absorb the data versatilely. Further, whereas searching images of a cat using a search engine would return numerous images of a cat as a result, the actual sample data quantity is not that large when it comes to trading.
Longine: In that case, would you say that the key is simply how you can apply knowledge of deep learning, rather than image recognition, to the trading system by making a slight customization?
Yokokawa: Yes. It would involve continuous customization. No academic paper with notable ideas is found, as nobody has done this. Although quite a few papers exist on image recognition, the necessary trading data is limited. Therefore, we learn by trial and error, as you would with street fighting, while asking ourselves, “What can we do?”
The Business Model of Capitalico
Longine: Going forward, how do you plan to shape Capitalico’s business model?
Yokokawa: Since Capitalico will be sold as a tool, one of the models we are considering is SaaS (Software as a Service), in which we launch it as a “freemium” and make money on the premium by setting points somewhere to charge fees. Another model involves charging a royalty to those who sell what they created. We are primarily considering these two.
Longine: Is there a company you compete with, or regard as a benchmark?
Yokokawa: In the sense of automating their own trade ideas, our rival is Quantopian in Boston. There is also a Russian company that sells a meta-software called MetaTrader. This is a huge company with a substantial market share in currency exchange system tools. There are other companies here and there, but their tools often make users perform programming-like tasks by drag and drop techniques, even though they should require no programming.
Rewards in Speaking and Receiving the Award at the FIBC2016
Longine: What was rewarding for you in attending the FIBC?
Yokokawa: I’m really happy that we won the award. First of all, it increased our exposure, as a lot of the media covered us. Also, the fact that investors recognize us because of the award is encouraging. I think that is the greatest reward.
I was also able to exchange thoughts with Meg Nakamura from Shift Payments overseas. I thought meeting interesting people overseas was really great.
Longine: Thank you for talking with us for such a long time today.
Yokokawa: Thank you very much.