What happened at SVB

Here’s my very simplified version of what happened to bring down SVB: and specifically, why no-one seemed to have seen it coming. Informed very much by Matt Levine’s excellent writing on the topic. Thoughts are purely my own, not representing any organisation.

At core, we need to look at a simplified model of what banks like SVB do, and especially, what then happens when interest rates change. Essentially, banks:

  1. Take in deposits, from individuals and businesses. Deposits are generally low-interest, and as interest rates rise, only a little of that is passed on to deposit accounts.
  2. Put all that money somewhere! Broadly, there are two options:
    • A. Loan-like instruments (e.g., home loans, business loans). These are often floating-rate, i.e., their interest rates follow market rates, but they are also very “illiquid” (hard to sell or otherwise turn into cash). If a bank makes a home loan for a specific house, it can’t easily get that money back immediately.
    • B. Bond-like instruments, like corporate debt. These are typically fixed interest rate, but they are liquid (easy to sell).

Now, what happens when interest rates go up? Deposits and bonds rates don’t really change much, but loan interest rates rise. This is an “endowment effect” that leads banks, all else being stable, to make more money when interests rise: their “Net Interest Income” (NII) rises as interest rates rise.

Great! Next question, what happens if, for some reason, a lot of depositors want their money back at once? The bank would eventually run out of cash reserves, and need to sell some bonds (as the loans are hard to sell). But here’s a problem: bonds have a fixed interest rate, but their market value decreases when interest rates rise, because new investors would rather buy new bonds offering a higher rate, than your old low-rate bonds. When a bank holds a bond to maturity, that’s not a problem — they get back the full face value of the bond. But, if a bond needs to be sold early, and the interest rates have risen, the seller will take a loss. At worst case, a bank being forced to sell lots of bonds could make a huge loss, which overwhelms its capital reserves, and leaves it insolvent.

Normally, this is irrelevant, as this only happens if a bank has to sell bonds early, i.e., has a massive outflow of deposits, a bank run. There are many mechanisms to prevent this:

  • deep relationships between the bank and it’s customers;
  • a wide variety of depositors, many of whom don’t really follow the finer points of financial news and so are fairly “sticky”;
  • deposit insurance;
  • capital buffers, regulatory supervision, risk modelling, etc etc.;
  • and hedges. Let’s talk about these.

Clearly, it would conceptually be useful for banks to be able to deploy cash in instruments that have both floating interest rates (and so do not lose market value when interest rates rise), and also highly liquid. You could imagine two ways to do that:

  1. Make loans more liquid, by, let’s say, packaging groups of similar loans into standardised instruments (call them “CDOs”), splitting them into tranches by risk, getting ratings agencies to rate them, and then create a liquid market for them. There’s a problem with though: it removes the risk from the loan originators, leading to perverse incentives that lead to bad quality loans, and you get the 2008 financial crisis. So, let’s not do this.
  2. Make bonds that don’t lose market value when interest rates rise. This can, broadly, be done by banks through hedging on interest rates. Then, when interest rates rise, the bonds lose market value, but the hedges make money to roughly counteract that effect, and vice-verse. This is a great idea, in general!

So why did SVB not have hedges in place? It seems that they were worried about what happens when interest rates fall: if hedges make money when rates rise, they obviously lose money when rates fall. Combined with the negative endowment effect on loans, this can make falling rates pretty bad for bank profitability. So, it seems that SVB dismantled much of their hedging in 2022, to take profits and to avoid losses if/when rates fell again. And this would have been fine, as long as we didn’t get both a rise in interest rates and a lot of depositors wanting their money back. Of course, that’s then what happened, and clearly the bank’s risk scenario testing was insufficient.

So let’s put this together into what led to SVB’s collapse:

  1. An (unrealised, theoretical) mark-to-market loss on bond holdings, due to:
    • lots of bonds relative to loans, at SVB, due to their client base of startups being relatively cash-rich and loan-light
    • insufficient hedging, due to concerns of the impact of hedges on profitability if rates were to fall.
  2. An unprecedented drop in deposits, due to:
    • a depositor base suddenly becoming less cash-rich, due to the sudden slowdown in VC funding to startups
    • a depositor base unusually prone to runs, because most of it was in deposits that exceeded the deposit insurance maximums, and came from depositors that were NOT diverse, as most startups (and especially their VC shareholders) were on the same Whatsapp groups
    • modern banking apps making it way easier to move cash out of a bank — no more queueing on the steps of the bank
    • some communication accidents and mistakes that flagged the theoretical massive losses on the bank’s bond holdings at market price.
  3. An inability to find extra liquidity to cover the gap:
    • SVB tried to raise further equity, but this failed and just contributed to the communication of the point just above, i.e., accelerated the deposit flight
    • emergency funding from the Fed, backed by bond holdings, would have had to have been done at market prices for bonds, thereby realising the theoretical mark-to-market losses, and leading to insolvency. Catch-22!

So my guess is, we’ll see regulatory changes and/or focus on requiring banks to model the impact of interest rate changes, not only on profitability and cash flow, but also on a bank’s ability to liquidate assets at short notice, without taking prohibitive market price losses.

Easily summarise PDFs (with AI)

Next chapter in learning and deploying AI tools: a tool to summarise arbitrary-length PDFs. Very useful for dry, long reports, legal judgements, etc. You could use it on a book, but it will be rather … dry?

Code at https://github.com/langabi/RecursiveSummarizer. It’s a simple (and not particularly elegant) Python script, but also a MacOS finder shortcut (right click on any PDF to summarise it, like a super-power on my computer!) Needs minor technical skill, and a (free) OpenAI.com API key.

The main point of this is to explore how to work around one of the key constraints of current Large Language Models like GPT-3: the limited size of the prompt. GPT-3 current version can, at time of writing, handle around 4000 tokens, or around 3000 words in the combined input and output. So, if you have a document bigger than this, chunking and recursive summaries are one way to go. It produces a decent result, at the risk of the summary losing a little of the overarching theme of a document.

More generally, a key technique to avoid AI models from, well, making things up, is to enrich the prompt with information relevant for an accurate answer to the prompt question. The limited prompt size is a constraint here too — you might find yourself using AI techniques like embedding to find exactly and only the needed content.

But certainly, I expect that there will be a lot of innovation in the next few months in how to give AIs working memory (long and short term), through things like architectural changes (bigger prompt windows), embedding-driven long-term memory, and maybe using the bots themselves to create a summary of a current conversation, that is repeated at the start of each new prompt window. And no doubt much more!

AI: The “spreadsheet” for copywriting

We just reduced new product copywriting time from 40 minutes to 15 minutes, with our first internal AI-powered tool. Here’s how I’m thinking about AI as the “spreadsheet for copywriting”.

Technically, the implementation relies on:

  • Enriching the “prompt” by replacing product- and order-specific references with all the information we have available, before passing the (much bigger) prompt to GPT-3, so that the AI is working on complete, current product and order information. Currently we’re just matching product/order IDs, but a more complete implementation will also used named entity extraction to find products, and then traditional search or embedding-based search to match and retrieve the matching products
  • Adding some boilerplate to the prompt about tone of voice, and instructing it to respond “I don’t know” when it is not confident in being able to answer the prompt.

Without this, we found (as have many others) that the AI makes up very plausible but completely fictional details.

For now, we’ve released the tool in an internal “chat bot”-style interface, but are looking at a Google Sheets function, to allow generating text for each of a list of products in response to a standard prompt — very useful for newsletter creation. Regardless, human review and editing is critical for all our use cases.

So far, we’re seeing fastest adoption by our copywriters, especially for listing new products. Faithful to Nature’s team, for example, spends hours reviewing each and every ingredient with our suppliers, before listing a product. This isn’t changing, but the process to convert that information into a standard, compelling product description, has dropped from 40 minutes to 15 minutes per product.

So where to from here? (I’m asking, and so, to be honest, are our copywriters).

I keep returning to a podcast by Planet Money on the impact of the spreadsheet on finance and accounting. Tl;dr:

  • spreadsheets used to be physical paper, calculated with hand calculators, laboriously
  • digital spreadsheets (Lotus 1-2-3, Excel) completely changed this, and away went all that human work
  • but, since then, finance employment has grown substantially
  • because spreadsheets changed from a reporting tool, to a scenario planning tool: what would happen if we opened another store? Changed our margin? So finance itself become far more strategic.

This is how I’m thinking about these AI copywriting tools: by replacing the “hand-held calculator” of having to write every single word by hand, we will achieve two things:

  • Unlock opportunity for copy, that was previously marginal. Ideally we’d have a page on our site for every possible combination of interests and questions for every small segment of customers, but that’s a lot of work. If the work halves, a lot more become possible — and indeed, we’re already working on newly-opened opportunities.
  • Make copywriting a strategic “what if” function. Currently, the marketing flow is linear: come up with a campaign concept, choose a tone of voice, identify copy requirements, write the copy, send to customer. This can become iterative: if 90% usable copy is generated instantly, we can experiment with different concepts based on how they turn out, before finalising the brief.

Clearly, elevating copywriting from  execution to what-if strategy is very applicable elsewhere, for example, in the legal industry.

If you’ve deployed AI tools yet, what have your experiences been?