FinTwin 2.0: what survived an out-of-sample test on 5,000+ companies

[SpiliosDimakopoulos]

FinTwin 2.0 is out: release notes · live dashboard · paper

This version tests the twin on every US non-financial SEC filer with $50M+ revenue (5,220 companies, including those that went bankrupt or were delisted), out of sample and point in time. The short version:

• On its own it is less accurate than a 16-quarter cash-flow margin. Most of what it seemed to know was persistence.
• The direction of its departures from that margin is informative. For free cash flow, blending in about 15% of them lowers out-of-sample error by more than any simple rule does.
• For bankruptcy ranking it does no better than its own scoring formula without simulation (AUC 0.86 vs 0.87).
• It over-reacted to revenue because it assumed too much fixed cost. A walk-forward-selected prior fixes much of that and is now the default.

Questions, criticism and ideas are welcome here: what to test next, companies or scenarios you'd want in the stress lab, or places where you think the method is wrong.