AI, Work and the Economy

Our view/Point 5/Indicator

Bank lending to non-bank lenders

NormalDownturn signalData to 11 Nov 2026

Loans were $2.14 trillion in the week of 11 Nov 2026, 4.8% higher than 13 weeks earlier, 19% higher than a year earlier and 3.9 times their 2019 average.

The change over 13 weeks is higher than 73% of the readings since 2015.

The prediction

OpenSet 13 Nov 2026Counts for point 5

  1. If we are wrongtriggers, then stops
  2. Now4.8%11 Nov 2026
  3. Start4.7%28 Oct 2026
  4. If we are right−3.0% or belowby 29 Sep 2027

If our view is right, this signal triggers by 29 Sep 2027: the change in US banks' loans to non-bank lenders over 13 weeks reaches −3.0% or below. The date comes from the model, in which the US unemployment signal triggers in July 2027. In the downturns since 1990 in which both triggered, this measure triggered at most two months after that signal, so the date is the end of the quarter that falls in.

If it triggers and then stops being triggered for four consecutive weeks, that counts against our view, in which the downturn, once begun, lasts well over a year. The deadline refers to the period the data cover. If it passes without a reading of −3.0% or below, that counts somewhat against our view.

Loans from US commercial banks to non-bank financial firms, $ billion
Change over 13 weeks, %

US recessionsPeriods when this signal was onThe level at which it triggers

What it measures
Loans from US commercial banks to financial firms that are not banks: private credit funds, business development companies, mortgage lenders and others (Federal Reserve, weekly).
Why it matters
Much of the lending for AI investment comes from private credit funds, which in turn borrow from banks. If banks cut these loans, losses on AI lending would spread to the rest of the financial system faster.
When it triggers
Triggers when the total falls 3% or more in 13 weeks.
How this signal counts: its place in the argument, its rule and its record

Where it fits

AI investment and how it is financed. The argument predicts: The build-out costs more than its builders earn and is financed at high long-term interest rates; lenders to its weaker borrowers pull back, and imports of computer hardware stop growing.

Other possible causes: The usual cycle in memory chips or in business investment; long-term interest rates that rise for reasons unrelated to AI, such as public deficits.

A trigger can count as evidence for point 5 of our view, a downturn about the size of 2001's.

When the rule was set

The rule was set on 25 Sep 2026, when it was not triggered. Its threshold is chosen by the author. In ten years of data it has triggered once, so there is little past data to judge its threshold against.

How often it has been on

Applied to past data at the end of each month from Mar 2016, using only what had been published by then, it was on in 0% of months to 2019 and 2% since 2020. It triggered once, within six months of a US recession dated by the National Bureau of Economic Research. The rule was set with these years in view, so these figures cannot show how well it predicts.

Past triggers

On the data published at the time, it triggered in Jul 2020.

Source: Federal Reserve Board (H.8) via FRED.