Skip to main content

Quant Analytics: from scratch to pro · Factors, volatility, regimes and lead-lag · Advanced

Market regimes without hindsight

Markets seem to switch between calm and turbulent states. A regime model describes those states with transition probabilities, but labels drawn afterwards use information nobody had at the time. This lesson separates regimes you can know in real time from those you only see in hindsight.

4 min read Reviewed Revision 1

In this lesson

  • Compute the expected duration and long-run share of each state from transition probabilities
  • Update the probability of a state with Bayes' rule after a new return
  • Tell a filtered (real-time) regime estimate from a smoothed (hindsight) one
  • Check that a regime rule uses only past data

Read first: Volatility models: EWMA and GARCH, Time series basics: random walks, autocorrelation and stationarity

Worked numbers

Each result is recomputed from the formula shown.

WhatFormulaResult
expected duration of the calm state, days1 / (1 - 0.98)50
expected duration of the turbulent state, days1 / (1 - 0.95)20
long-run share of calm days(1 - 0.95) / ((1 - 0.98) + (1 - 0.95))71.4%
long-run share of turbulent days1 - share_calm28.6%
unconditional daily volatility of the mixsqrt(share_calm * 0.006**2 + share_turb * 0.018**2)1.088%
calm density at -3%exp(-0.5 * (0.03 / 0.006)**2) / 0.0060.00062
turbulent density at -3%exp(-0.5 * (0.03 / 0.018)**2) / 0.01813.85
probability of turbulent after a -3% day0.02 * lt / (0.02 * lt + 0.98 * lc)99.78%
calm density at -1.2%exp(-0.5 * (0.012 / 0.006)**2) / 0.00622.56
turbulent density at -1.2%exp(-0.5 * (0.012 / 0.018)**2) / 0.01844.5
probability of turbulent after a -1.2% day0.02 * lt2 / (0.02 * lt2 + 0.98 * lc2)3.87%
average age of data in a 20-day window(20 - 1) / 29.5

Check yourself

A state persists from one day to the next with probability 0.95. What is its expected duration?

1 / (1 - 0.95) = 20 days.

Calm persists with probability 0.98 and turbulent with 0.95. What share of days is turbulent in the long run?

The calm-to-turbulent chance is 0.02 and the turbulent-to-calm chance is 0.05. The turbulent share is 0.02 / (0.02 + 0.05) = 28.6%.

A chart labels March 2020 as a crisis regime from the start. Why can that label not be used to trade on 5 March?

The start date was known only after later data confirmed the episode. On 5 March the label did not yet exist. Trading on it uses the future.

A 20-day trailing volatility is used as a regime signal. How old is the information in it on average?

(20 - 1) / 2 = 9.5 days. A trailing rule always reports a regime late.

A model finds three regimes in 10 years of data. Does that prove three regimes exist?

No. The number of states is a modelling choice. A sample with a few real episodes cannot separate three states from two with noise.

Key terms

Market regimeHalf-lifeExponentially weighted moving averageGARCH

Checked facts

Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).

Sources

Informational content. Not investment advice.