We provide large-scale evidence on the occurrence and value of FinTech innovation. Using a unique dataset of patent filings covering 2003-2017, we apply text-based machine learning to identify and classify innovations according to their key underlying technologies. To measure the value of FinTech innovations for firms and industries, we develop a new method that combines stock price responses with estimated patent-filing count intensities.
Our cross-sectional analysis shows that most types of FinTech innovation yield positive value to innovators, with Blockchain being particularly valuable. For the financial sector as a whole, Internet of Things (IoT), Robo-Advising, and Blockchain are the most valuable innovation types. Innovations impact industries more negatively when they involve disruptive technologies that originate from nonfinancial startups. Also, market leaders that have invested heavily in their own innovation appear to avoid much of the negative value effect from disruptive innovation by startups.
We then apply several families of machine-learning algorithms to the textual data to identify FinTech innovations and classify them into seven key technology categories: Cybersecurity, Mobile Transactions, Data Analytics, Blockchain, Peer-to-Peer, Robo-Advising, and Internet of Things.
Based on a general reading of various articles and reports, we formulate a broad typology of FinTech. First, we use text-based filtering to narrow down the large set of Class G&H patent filings to those that are plausibly related to financial services. Second, we apply several families of machine-learning algorithms to automatically generate classifications of the filings.
We construct a new lexicon of financial terms that can be used to exclude nonfinancial patent filings from the overall sample.
First, we obtain all terms from Campbell R. Harvey’s Hypertextual Finance Glossary (November 8, 2016 version), a widely-used online compendium of finance terms that serves as the basis for the glossaries of numerous media companies such as New York Times, Forbes, CNN Money, and others.
Second, we gather all terms from the online Oxford Dictionary of Finance and Banking, 5th Edition, published by Oxford University Press. Combining these two lists of terms and excluding acronyms, we obtain a total of 11,196 unique single-word and multi-word finance terms.
We review and manually classify the 1,000 filings into nine groups (seven for the FinTech categories, one for other financial filings, and one for nonfinancial filings). Using these groups as the basis for a simple nearest-centroid classifier, we classify the entire sample of 67,948 text-filtered filings into the nine groups and then select 200 from each group. The resulting collection of 1,800 filings, which we manually reclassify as needed, constitutes our training sample.
Support Vector Machines (SVM) and Neural Networks.
To explore the implications of FinTech innovation, we develop a new methodology for estimating the value of a patent filing to one or more publicly-traded firms. Our valuation approach is based on observed stock market responses to USPTO disclosures of patent filings. Importantly, our approach accounts for the market’s anticipation of different types of patent filings made by different filers. We start by estimating innovation arrival intensities via Poisson regression models that account for factors such as technology type, time effects, a patent filer’s prior innovation experience, and filer fixed effects. Then, for each patent filing, we combine its predicted count intensity with a firm’s stock price movements to infer the underlying value of the innovation to the firm.