1.8.6 Advanced Extensions

Basic Prerequisites: 04 · Hidden Markov Models and Mean–Variance (or Risk Parity).

---

1. Intuition & Practical Objective

Regime detection is not an end in itself - its payoff is deciding what to do differently given the regime you are probably in. This page extends the models three ways: (1) regime-based tactical allocation (shift equity weight with the filtered bull probability), (2) richer regime dynamics (regime-dependent volatility - Tsay Ch 12’s Markov-switching GARCH-M; multivariate regimes), and (3) the Bayesian/MCMC estimation frontier (Gibbs, Metropolis–Hastings, FFBS) as the alternative to EM.

The practical claim (Ang & Timmermann 2012; Kritzman, Page & Turkington 2012): returns in different regimes have different means, vols, and correlations, so a static allocation is wrong in every regime by construction - it is a compromise tuned to no state. A regime-aware allocation tilts toward risky assets when the bull probability is high and de-risks when it collapses, targeting the conditional opportunity instead of the unconditional one.

---

2. Mathematical Ground Truth & Derivations

Regime-conditional risk. In a 2-regime world the unconditional return distribution is the regime-weighted mixture

$$f(y_t)={\pi }_{1}N({\mu }_1,{\sigma }_1^2)+{\pi }_{2}N({\mu }_2,{\sigma }_2^2),$$

so the unconditional variance mixes the two volatilities and the squared mean gap:

$$\mathrm{Var}(y)={\pi }_1{\sigma }_1^2+{\pi }_2{\sigma }_2^2+{\pi }_1{\pi }_2({\mu }_1-{\mu }_2{)}^2.$$

The last term is the volatility of the regime itself - the reason regime models explain fat tails and vol clustering that a single Gaussian cannot (Ang & Timmermann’s survey frames this as the economics of regime switches).

Allocation policy. A natural regime-aware rule: hold equity weight $w_{\text{high}}$ when the filtered bull probability exceeds a threshold, $w_{\text{low}}$ otherwise,

$$w_t=\{\begin{array}{ll}{w}_{\text{high}}& {\hat{\xi }}_{t\mid t,1}>\tau \\ w_{\text{low}}& \text{otherwise}\end{array}$$

using the one-period-lagged filtered probability (only causal information). This is exactly the Ang–Timmermann portfolio-choice setting: the conditional mean/variance used in a mean-variance (or risk-parity) objective is regime-weighted, so the optimizer’s inputs - and hence its weights - move with the filtered state.

Estimation frontier - MCMC (Tsay Ch 12, verified). EM gives point estimates; the Bayesian alternative samples the full posterior. The pieces (all verified against tsay_ch10-12.md):

• Gibbs sampling: iterate draws of each parameter from its full conditional given the others and the data; discard burn-in $m$, treat the rest as an approximate iid posterior sample. Point estimate ${\bar{\theta }}_i=\frac{1}{n-m}{\sum }_{j=m+1}^n{\theta }_{i,j}$.
• Metropolis–Hastings: for intractable conditionals, accept a candidate with probability $\min (1,\frac{f({\theta }^{\ast })J({\theta }_t\mid {\theta }^{\ast })}{f({\theta }_t)J({\theta }^{\ast }\mid {\theta }_t)})$.
• FFBS (forward-filtering–backward-sampling, Ch 12.8): run the forward filter, then draw the whole latent state sequence jointly backward - the natural sampler for regime/state-space latent variables.
• Markov-switching GARCH-M (Ch 12.9): regimes drive both the mean and the conditional volatility, with the GARCH coefficients themselves regime-dependent; estimated via Gibbs with Griddy-Gibbs on the nonlinear coefficients.

---

3. Computational Implementation - regime-based tactical allocation vs static

Simulate a 2-regime equity market (bull: mean $+1.6\%$, vol $2\%$; bear: mean $-2.0\%$, vol $4.5\%$) plus a low-risk bond, run a Hamilton filter to get the live bull probability, and compare a static 60/40 against a regime-tactical rule (95% equity when $P(\text{bull})>0.6$, 20% otherwise). Stdlib only.

On this synthetic path the regime-aware rule raises the Sharpe by ~70% ($0.67$ vs $0.39$) and cuts the worst drawdown by about three-quarters ($-5.7\%$ vs $-24.2\%$) by cutting equity weight the $103$ periods the filter judged bearish. (The wealth figures are compounding artifacts of the synthetic per-period returns - compare the Sharpe and drawdown, not the terminal wealth.) This is the concrete payoff of regime detection: it converts a static compromise into a state-conditional bet.

---

4. Failure Modes & First-Principles Breakdowns

1. The allocation inherits the detection lag. You trade on the filtered (causal) probability, which lags the true regime - exactly the cost that quantifies how regime-aware allocation must be. The $88.6\%$ smoothing accuracy of 04 · HMM is not available to a live trader; the causal filter is weaker.
2. Parameter estimates are load-bearing. The allocation trusts ${\mu }_j,{\sigma }_j,P_{ij}$ estimated under a specific regime count and identification. Feed it a label-switched or over-fit model (05 · Failure Modes) and the “bull/bear” tilt is noise. Penalize (BIC) and constrain (ordering) before allocating.
3. Turnover / whipsaw cost. Switching weights with a noisy filtered probability churns the portfolio and pays transaction costs; the tactical rule above ignores these. A persistence filter (only switch after $\tau$ stays above/below the threshold) trades latency for turnover.
4. Regime homogeneity is an assumption. A “bull” equity regime may coexist with different bond/credit/correlation behavior; univariate regime detection says nothing about the joint structure. Multivariate regime models (Tsay Ch 10) are needed for that.

---

5. References

• Ang, Andrew & Timmermann, Allan: Regime Changes and Financial Markets, Annual Review of Financial Economics 4, 313–337 (2012)
• Kritzman, Mark, Page, Sébastien & Turkington, David: Regime Shifts: Implications for Dynamic Strategies, Financial Analysts Journal 68(3) (2012)
• Tsay
• Hamilton (1989)

---

6. Connected Graph Bridges

• Back: 04 · HMM · 05 · Failure Modes · Index Hub
• Portfolio: Risk Parity (vol targeting) · Mean–Variance · Black–Litterman (views meet regime estimates)
• Risk: ES) · Stress Testing & Scenario Analysis (regimes as scenarios)
• Bayesian/MCMC base: Bayesian Statistics · ML cross-link: Regime Classification: HMM & GMM