Alpaca Japan, a capital-markets fintech that has developed a roster of bank and broker-dealer clients and pilots, is using resources from a recent fund raise and a partnership to begin marketing to the buy side.
Morifumi Yotsumoto, Tokyo-based co-founder and CEO, met with DigFin in Hong Kong during a regional tour to pitch global hedge funds and other buy-side institutions. “We’re now getting requests from global investors,” he said.
Alpaca uses machine learning to predict short-term (30 minute) directions of major currency pairs, as well as other asset classes such as equities.
The company’s origins are in San Francisco, where a Japanese ex-banker, Yoshi Yokokawa, set up Alpaca Markets, which has developed a consumer robo business (including crypto). In 2016 it set up a Tokyo branch, but for domestic B2B sales. Yotsumoto joined the Japan business as CFO.
(In one of those strange twists, “Yotsu” had been the one who hired Yokokawa, then a fresh grad, into the investment-banking world in Japan, when Yotsumoto was a debt-capital markets executive at Lehman Brothers.)
Last year, Yotsumoto and the Japan team bought out the domestic business (Yokokawa remains a minority shareholder). Both arms of Alpaca had been looking to raise capital, but the U.S. branch found it hard to convince American venture capitalists to put money into a Japanese B2B fintech; and the Tokyo team sensed it could do more if they controlled their destiny.
The B2B Japan business developed its own A.I. tech for market making, especially in foreign exchange, which it provided to broker-dealers’ quant teams to help them gain an edge.
(Specifically, Alpaca specializes in convolutional neural networks, or CNNs, a field of A.I. that enables visualization of patterns involving many correlations. Alpaca crunches data across dozens of market factors to look for trends involving low-latency bid/offer spreads, mid-price moves, and tick density. The overall effect is like a chartist providing intense bursts of insights.)
To scale globally we couldn't just hire more people
The business used the client’s data to train its algos within a proprietary data warehouse, build APIs into the client’s systems, and run prediction signals. Alpaca competes globally with the likes of Two Sigma, and with investment banks' proprietary tech.
As some clients such as Bank of Tokyo-Mitsubishi grew comfortable with the service, Alpaca could tweak its models and add correlations with more asset classes – with results taking anywhere from six to 18 months to materialize, depending on the size of the client’s data team. (It takes a lot of people to train the machine.)
New business models
Then, starting in July 2018, the Japan business forged a partnership with Bloomberg. “If we were to scale globally, we couldn’t just hire more people,” Yotsumoto said. “We needed a distribution partner.” Bloomberg’s salespeople will help sell an Alpaca app, with a revenue split. And Alpaca can now source Bloomberg’s real-time data to develop products that don’t rely on clients. (That’s for the app – not for bespoke algos Alpaca sells directly: for those it must purchase the data from either Bloomberg or another source.)
Yotsumoto says the app that Bloomberg can sell won’t generate a lot of clients, at least not yet: it’s basic, just a 30-minute predictor of some forex pairs and equity indices. But it’s already leading to new connections, particularly among banks in Singapore and Taiwan.
That, plus a recent fund raise, is giving Alpaca the chance to now sell its software to buy sides, especially American hedge funds.
Alpaca closed a $6 million Series A round of financing in September (along with a $1 million investment from Japan Finance Corporation, a quasi-sovereign lender). The round valued Alpaca Japan at $45 million, at a 10x earnings multiple.
Alpaca’s algos make predictions for time frames ranging from one minute to one hour, it is now developing those for two weeks or one month. Buy sides want even longer time frames, so Alpaca is exploring using economic data (inflation or consumption information, for example) as well as market data. But this is new and tests are not yet conclusive as to whether this provides a reliable edge.