The regime your data never contained

Graduate · enrolled learners

This lesson opens with Everest Capital and the Swiss franc, 15 January 2015.

What happened

For years the Swiss National Bank held a floor under the euro-franc exchange rate, promising to keep the franc from strengthening past a set level, and that floor had held so consistently that it looked like a fixed feature of the market. Everest Capital's flagship Global fund carried a position that implicitly assumed the floor would remain, a reasonable bet given every day of recent history. On 15 January 2015 the Swiss National Bank abruptly abandoned the floor, and the franc surged by roughly thirty percent in minutes, a move that in the world the fund's models had known was essentially impossible. The fund lost essentially everything, around 830 million dollars, and closed within days. Nothing was miscalculated in the ordinary sense: the models and the record were built on a period in which the floor always held, so they had never seen the floor break, and a strategy validated only against a regime that persisted throughout its data is a strategy that has never been tested against that regime ending. The event that destroyed the fund was not in the history the fund had learned from.

The decision point

A model or a backtest learns from the regime that was present in its data, the stable relationships and boundaries that held throughout the sample, and it is blind to those relationships breaking, because a break that never happened in the history cannot be in the record, so a strategy can be thoroughly validated against a regime and completely untested against that regime ending. Everest Capital is the case: a bet that implicitly relied on the Swiss franc floor was validated by years in which the floor always held, and when the Swiss National Bank abandoned it and the franc jumped thirty percent in minutes, the fund lost essentially everything, around 830 million dollars, because the event that ruined it was one its data had never contained. This extends the course's theme from a missing tail to a missing regime: it is not only rare events within the same regime that a backtest can omit, but the change of regime itself, the peg that breaks, the correlation that inverts, the rule of the game that is rewritten, and these are precisely the events that produce the largest losses and are least likely to appear in a sample drawn from the stable period. The error is to treat a relationship that held throughout the data as a fixed feature of the world, when it may be a temporary regime that can end. So the discipline is to identify the regime a strategy depends on, the pegs, bounds, correlations, and structures it assumes will hold, and to stress-test the strategy against those assumptions breaking even though the data never showed them break, because Everest's models were right about every day they had seen and silent about the one day that mattered, which was the day the regime changed.

Recorded loss

$830,000,000

Orientation and Year One are open: anyone can read them without an account. From Year Two onward the lessons are for enrolled learners, because progress through the later years only means anything if it is tracked against a record.

It is free. We do not sell the list and there is nothing to buy at the end of it.