LTCM’s risk model deliberately used conservative estimates above observed values.

Evidence

The standard deviations and correlations that went into LTCM’s aggregaterisk model were, however, not simply the empirically observed numbers; they were deliberately conservative estimates of future values. The observed standard deviation of the U.S. dollar swap spread, for example, was around 12 basis points a year, while, as noted above, the risk model assumed it would be 15 (Rosenfeld interview). Past correlation levels, likewise, were “upped” to provide a safety factor: despite observed correlations being 0.1 or less, LTCM was “running analyses at correlations at around 0.3” (Meriwether interview). — Donald MacKenzie, An Engine, Not a Camera, 998

Categorization

• XREF: Reverses the common post-crash narrative that LTCM’s models were recklessly overconfident — its risk parameters were actually conservative relative to history. Speaks to how a properly risk-managed model can still blow up via regime change. — connects to: LTCM’s own modelers tempered their risk model with judgment-based safety factors. — Donald MacKenzie, An Engine, Not a Camera: How Financial…, 1018; LTCM’s crisis did not stem from blind faith in mathematical models. — Donald MacKenzie, An Engine, Not a Camera: How Financial…, 1017
• CANDY: The detail that LTCM deliberately upped correlations to 0.3 as a safety factor, and it still wasn’t enough, is a striking conversation-stopper about model limits.

Broader Connections

• LTCM’s own modelers tempered their risk model with judgment-based safety factors. #expands (0.92) — Donald MacKenzie, An Engine, Not a Camera, 1018
• LTCM’s crisis did not stem from blind faith in mathematical models. #contradicts (0.78) — Donald MacKenzie, An Engine, Not a Camera, 1017