How autoML is breaking into investment banking

At the core of an investment bank’s work is valuing companies. A buyer and a seller, for example, will have to agree on a valuation to trade financial assets or equities such as a company’s stocks or bonds.

Without the work of investment banks, much of the exit magic that happens in the tech industry wouldn’t take place, firms wouldn’t be able to trade on stock markets, and we wouldn’t know how much our shares are worth.

However, Tommy Tan, CEO of boutique investment bank TC Capital, says many of the valuation methods used by bankers today “are deeply flawed.”

Traditional methods are outdated

Traditionally, investment bankers in Wall Street and all over the world value firms through three standard methodologies.

The first two methods involve comparing past mergers and acquisitions (M&A) or stock market valuations of similar companies. The third way uses discounted cash flow (DCF) models to help bankers estimate the present value of a company by projecting the stream of cash flows that a company being bought or sold can generate, and then discounting its total cash flow.

But these techniques are manually intensive and carry a high risk of human error. Worst of all, the valuations they produce can be highly subjective. Data from comparable M&A deals, for instance, could be spread over several years, rendering such information irrelevant. In the trading comparables method, it’s difficult to find publicly traded firms that are similar enough to the one being evaluated because every company is unique. In addition, granular data about financing in the fast-moving and high-growth startup industry “is scarce and no information source provides a satisfactory solution,” says Tan.

TC Capital CEO Tommy Tan / Photo credit: TC Capital

Instead of adding robustness, these qualitative considerations make a single valuation even more subjective. Using standard metrics like price-to-earnings-ratios, which refer to the attractiveness of a company’s stock price compared to its current earnings, can be unreliable because even comparable companies could have vastly different ratios. A company experiencing a bad year could trade at high ratios.

Finally, “the DCF method is beautiful theoretically, but highly subjective. Assumptions used to forecast future cash flows and the resulting valuation are highly sensitive to innocent-looking changes,” points out Tan.

For instance, even a half-percent change in the discount rate can swing valuations by 30%.

Say ‘hello’ to the Machine of Wall Street

Though TC Capital relied on these methodologies in 2002 when it was first founded, it now seeks to build its own valuation methodology with automated machine learning (autoML): the process of automating applied machine learning to real-world problems.

“We took the trading comparables approach, as it is entirely market-driven in transparent exchanges and high levels of disclosure. The data will be of reasonable quality,” says Tan.

With these data points, TC Capital built a trading comparable table that includes information from 43,000 companies and assembled 560 variables on each firm.

“It is humanly impossible to perceive patterns in this table, but that is what autoML is designed to do – analyze this data to find patterns that explain the different valuations of companies,” he explains.

It took TC Capital nine months to assemble a robust dataset composed of a proprietary mix of financial, macro, industry, and web-sourced data. Beyond venture capital reputations and track records, rich datasets would entail attributes of companies such as their growth rates of revenue, average cost of customer acquisition, gross margins, EBITDA margins, etc.

By using a machine learning model backed by a robust dataset, the company was able to generate rich features (individual, measurable attributes) and account for differences in company valuations sufficiently. Tan and his team were then convinced they were on the right track.

Eureka, it works!

To make the process painless, TC Capital worked with enterprise AI firm DataRobot to outsource its machine learning platform. “[DataRobot] took care of not only autoML, but also normalizing the data for [machine learning] and all the tedious housekeeping and sanity checks,” Tan shares.

Now, when voting as a member of TC Capital’s valuation committee, he relies on the valuation that the autoML model spits out.

The firm’s machine learning-generated valuation models consistently explain 90% of the differences in the valuation of companies.

“We’ve also backtested (testing a predictive model on historical data) the approach for 20 years, and the results were surprisingly good,” Tan claims. “With over 560 features, the data was rich enough for the machine learning to detect valuation patterns that are not possible with the traditional small table methods.”

The backtesting measured whether these trading strategies made money – and they did, beating the market by nearly 3x. Tan, who believed that beating the market was impossible, even audited the numbers three times, one of which was an independent audit.

The model is both consistent and highly accurate – characteristics that allow TC Capital’s bankers to stand by their valuations at the negotiating table.

An investment banker in an app for CEOs

AutoML can also be applied on a larger scale to help C-level executives value public and private companies. TC Capital’s app, CeeSuite, does exactly this, leveraging intelligence sources like machine learning, data science, crowd-sourced intelligence, and the knowledge of human experts to solve the problems that CEOs face, says Tan.

TC Capital’s ultimate and ambitious goal is to build the world’s first or biggest investment bank that doesn’t employ a single banker. The first step is to launch CeeSuite in January 2020.

If you think that this autoML application to banking sounds simple, that’s because it is. The investment banking world hasn’t changed much, compared with consumer-facing industries like entertainment and retail.

“It’s not easy to explain the autoML methodology in an industry that still uses 19th-century tables for valuation,” Tan says. “But we now have a truly modern approach to valuation which is objective, accurate, and immune to subjective choices by people preparing tables.”

DataRobot is the leader in enterprise AI, delivering trusted AI technology and ROI-enablement services to global enterprises competing in today’s Intelligence Revolution. Its enterprise AI platform maximizes business value by delivering AI at scale and continuously optimizing performance over time.

Read DataRobot’s ebook to learn more about the practical use cases of AI in today’s investment banking market.