Perspectives on Risk - Feb. 22, 2022 - FinTech (Non-crypto)
UBS & Wealthfront; Machine-Learning the Skill of Mutual Fund Managers; Incentives Matter (Paypal edition); FHFA Issues Guidance on AI/ML; Eventually, we’ll all be using ML in Excel;
I still haven’t finished reading and thinking about the CBDC, but here is a bunch of non-crypto observations.
UBS & Wealthfront
UBS’s announced in January that it plans to acquire Wealthfront for $1.4 billion.
Tommy Marshall, executive director of Georgia Fintech Academy, pointed out:
Wealthfront in many ways helped start the robo-advising trend when it was founded over 13 years ago.
This is UBS’s third dip into robo advice. UBS first launched Smart Wealth, a European-based offering, in 2017 only to shutter the platform in 2018. In 2018, nearly two years after making an investment in SigFig, which builds robo advisors for large financial institutions like Wells Fargo, UBS rolled out UBS Advice Advantage. It offered hybrid digital advice at a $10,000 minimum investment and 75 basis point management fee.
So is UBS buying the clients or the technology?
[T]he technology is easily replicable, as Schwab, Fidelity and Vanguard have dominated the robo space soon after Wealthfront entered.1”
There really isn’t much value in the client base with an average account size of ~$57k, meaning about $140 per year in revenues (25 bps) — nobody wants that.
Then there is the issue of client defections. Wealthfront had set out to “disrupt advisors and traditional wealth management via technology” but has now “sold the business to a 160-year-old bank and wirehouse — just the people they claimed were evil and had targeted to disrupt.”
“Wealthfront is so anti-establishment that there will be client defection for those who wanted an independent solution and not get swept up into a 160-year-old institution,”
“One of the attractive things about Wealthfront is that it was not a traditional, legacy advice firm,” Goldstone said. “Some clients may not be excited about becoming a UBS client overnight.”2
Machine-Learning the Skill of Mutual Fund Managers
We show, using machine learning, that fund characteristics can consistently differentiate high from low-performing mutual funds, as well as identify funds with net-of-fees abnormal returns. Fund momentum and fund flow are the most important predictors of future risk-adjusted fund performance, while characteristics of the stocks that funds hold are not predictive. Returns of predictive long-short portfolios are higher following a period of high sentiment or a good state of the macro-economy. Our estimation with neural networks enables us to uncover novel and substantial interaction effects between sentiment and both fund flow and fund momentum.
The model:
identifies fund characteristic information, and specifically fund flow and fund momentum, as the key predictors of mutual fund out-performance. Moreover, these two fund characteristics matter much more when investor sentiment is high. That is, there is an important interaction effect, which linear models fail to pick up.
Buying the ten percent of mutual funds with the best predicted performance each month, and using the model not only to select but also to weight the funds within the top decile, generates a cumulative abnormal return of 72% over our sample period. Buying the ten percent of mutual funds with the worst predicted performance each month produces a cumulative abnormal return of -119%. The 191% difference in out-of-sample performance based on the model’s predictions is economically large and statistically significant.
Pretty available data, pretty well-known modeling techniques. Welcome to the future bot wars.
Incentives Matter (Paypal edition)
PayPal Admits 4.5 Million Accounts Were Illegitimate As Fintech’s Fraud Problem Grows
PayPal chief financial officer John Rainey said the company identified 4.5 million accounts that it believes “were illegitimately created.”
Over the past two years, while ecommerce soared due to the pandemic, PayPal added 120 million new customers (it now has 426 million total accounts). In 2021, the company “leaned into incentivized customer acquisition tactics to a much greater extent than we ever have in our history,” Rainey said on the earnings call. For example, PayPal ran marketing campaigns that offered to deposit $5 or $10 in a new customer’s account if he or she signed up for PayPal or Venmo. It ran into trouble when bots, or software created to automatically visit websites and take actions, started scooping up those incentives for the sole purpose of seizing the reward.
“How much of the growth in some of these fintech portfolios could be fueled by these bots that are creating accounts just for collecting the incentive?” McKenna says. “I think every fintech should look at accounts that signed up with incentives but never used the account again.”
“What we’re seeing at PayPal is a systemic issue,” says Mary Ann Miller, a vice president at identity and fraud company Prove. “It’s related directly to the identity theft and synthetic fraud that we saw during the pandemic.” She says that bad actors are weaponizing the personal information that they’ve stolen in data breaches and using bots to launch attacks. “They’re going to all kinds of fintechs and attacking their account-openings processes,” she adds.
FHFA Issues Guidance on AI/ML
The first US agency has issued guidance on the use of machine learning; the FHFA has issued AB 2022-02 Artificial Intelligence/Machine Learning Risk Management that provides “guidance to Fannie Mae and Freddie Mac on managing risks associated with the use of artificial intelligence and machine learning (AI/ML).”
Those familiar with regulatory guidance will see all of the usual terms:
The sophistication of the AI/ML risk management activity should be proportionate to each Enterprise's size, complexity, and risk profile.
The Enterprise should leverage enterprise-wide risk management and control frameworks, including those used for model, data, technology, information security, third-party, and compliance risk management, to the extent practicable.
The degree and scope of risk management and controls addressing AI/ML should be risk-based and commensurate with the extent and complexity of AI/ML development and use at the Enterprise, as well as the level of risk exposure.
It requires Governance (roles & responsibilities, policies & procedures), Risk Identification and Assessment (definitions & taxonomy; inventory; model, data and operational risk considerations; fair lending & compliance), a Control Framework, and expectations around Risk Monitoring, Reporting, and Communication.
Importantly, it imposes expectations that you will understand third-party developed ML models.
A bigger, unaddressed issue for regulators will be to devise a regime that allows, and indeed requires, firms to look for hidden bias, Firms may choose willful blindness if attempts to detect these biases raise legal and/or regulatory risks. Firms will need a 'safe harbor' if they are to be encouraged to check.
Eventually, we’ll all be using ML in Excel
Should just anyone be given keys to the AI machine? Why low-code AI tools pose new risks. [BP: one of the better articles on the risks of unleashing these tools broadly]
Low- and no-code AI tools rely on visual interfaces with drag-and-drop functions and drop-down menus for building machine-learning models.
But even people who see the value in these tools worry that gifting amateurs with Easy-Bake AI Ovens is a recipe for risk.
Some AI practitioners themselves are leery of an onslaught of AI made by people who lack knowledge of standard processes for debugging as well as testing for quality control and reliability. They worry that people who aren’t trained on the nuances of how machine learning works could unintentionally unleash AI that makes discriminatory decisions.
K-Means is one of the most common unsupervised machine learning algorithms. In this article, I will implement one algorithm in Excel from scratch with a simple dataset to find the centroids.
As you may already notice, in a series of articles, I use Excel/Google Sheet to implement the basic machine learning algorithms so that we can understand the underlying principles:
Shit.
I Haven’t Gotten Through the CBDC Stuff Yet
Here are a bunch of recent Federal Reserve papers and things; I hope to comment once I’ve digested them.
Money and Payments: The U.S.Dollar in the Age of Digital Transformation (FR)
Project Hamilton Phase 1 Executive Summary (FRB-Boston)
Stablecoins: Growth Potential and Impact on Banking (FRBOG)
Preparing for the Financial System of the Future (FR Brainard speech)
A Primer on Central Bank Digital Currencies (Treasury)
Central bank digital currencies (Bank of England)
More Reading
Privacy is Dead