A market event can affect several assets differently, making it hard to understand the change from one price chart. An equity series may look much as it did before, while oil, gold or currencies tell another part of the story. Daily prices alone register changes around Lehman, COVID, the Ukraine invasion and Liberation Day.
Across five markets, the Bayesian detector finds 69 changepoints. Historical checks give 47 clear event matches, 10 plausible ones and 12 tenuous associations. We extend the detector to joint returns and check those matches, because a detected statistical change does not automatically identify its cause.

In the first post in this series, PELT found regime shifts in S&P 500 returns by processing the entire history at once. In the second post, we implemented univariate BOCD: an online algorithm that maintains a probability distribution over the current regime's age as each new data point arrives. Both methods fed on a single series: S&P 500 daily log returns.
This post is the direct follow-up promised in that second post. We noted there that the Iran war (28 February 2026) produced no detectable signal in S&P 500 returns alone. The market was already turbulent from Liberation Day tariffs, so the additional volatility did not register as a new regime. But oil futures spiked sharply, gold rose as a safe haven, emerging market equities fell on regional exposure, and the dollar strengthened. Any of those series, watched individually, might not be distinctive enough. Watched simultaneously, the cross-asset co-movement pattern is the signal.
By the end of this post, you will have extended BOCD to dimensions using a Normal-Inverse-Wishart (NIW) conjugate model, applied it to S&P 500, oil, gold, emerging market equities, and the dollar index, and seen why the covariance between asset classes carries regime information that no single series captures alone.
Quick Win
Start by downloading aligned daily log returns for all five assets:
import numpy as np
import pandas as pd
import yfinance as yf
from scipy.special import gammaln
TICKERS = {
"^GSPC": "S&P 500",
"CL=F": "Oil (WTI)",
"GC=F": "Gold",
"EEM": "EM Equities",
"DX-Y.NYB": "USD Index",
}
raw = yf.download(list(TICKERS.keys()), start="2003-10-01",
auto_adjust=True, progress=False)["Close"]
raw.columns = list(TICKERS.values())
# Align trading calendars and compute log returns
log_ret = np.log(raw / raw.shift(1)).dropna()
print(f"{len(log_ret)} aligned trading days, {log_ret.index[0].date()} to {log_ret.index[-1].date()}")
Now the two core components: the MultivariateStudentT class (which replaces StudentT from Post 45) and the bocd_multivariate function.
class MultivariateStudentT:
"""NIW conjugate model — predictive: multivariate Student-t."""
def __init__(self, d, mu0=None, kappa0=1.0, nu0=None, Psi0=None):
self.d = d
self.mu0 = mu0 if mu0 is not None else np.zeros(d)
self.kappa0 = kappa0
self.nu0 = nu0 if nu0 is not None else float(d + 2)
self.Psi0 = Psi0 if Psi0 is not None else np.eye(d) * 1e-4
# Sufficient statistics (n_rl rows, one per run-length hypothesis)
self.mu = self.mu0[None, :] # (1, d)
self.kappa = np.array([kappa0]) # (1,)
self.nu = np.array([self.nu0]) # (1,)
self.Psi = self.Psi0[None, :, :] # (1, d, d)
def log_pred_prob(self, x):
"""Log predictive probability under each run-length hypothesis."""
d = self.d
df = self.nu - d + 1 # (n_rl,)
sf = (self.kappa + 1) / (self.kappa * df) # (n_rl,)
S = self.Psi * sf[:, None, None] # (n_rl, d, d)
# Cholesky for log-det and quadratic form
S += np.eye(d) * 1e-10 # numerical jitter
L = np.linalg.cholesky(S) # (n_rl, d, d)
log_det = 2 * np.sum(np.log(np.diagonal(L, axis1=1, axis2=2)), axis=1)
diff = x - self.mu # (n_rl, d)
z = np.linalg.solve(L, diff[:, :, None]).squeeze(-1) # (n_rl, d)
quad = np.sum(z**2, axis=1) # (n_rl,)
return (
gammaln((df + d) / 2) - gammaln(df / 2)
- (d / 2) * np.log(df * np.pi)
- 0.5 * log_det
- ((df + d) / 2) * np.log(1 + quad / df)
)
def update(self, x):
"""NIW Bayesian update — prepend prior for run_length = 0."""
diff = x - self.mu # (n_rl, d)
new_kappa = self.kappa + 1
new_nu = self.nu + 1
new_mu = (self.kappa[:, None] * self.mu + x) / new_kappa[:, None]
outer = np.einsum("ni,nj->nij", diff, diff)
new_Psi = self.Psi + (self.kappa / new_kappa)[:, None, None] * outer
self.mu = np.concatenate([self.mu0[None, :], new_mu ], axis=0)
self.kappa = np.concatenate([[self.kappa0], new_kappa])
self.nu = np.concatenate([[self.nu0], new_nu ])
self.Psi = np.concatenate([self.Psi0[None, :, :], new_Psi ], axis=0)
def bocd_multivariate(data, hazard_rate=500, kappa0=1.0, nu0=None, Psi0=None):
"""Multivariate BOCD (Adams & MacKay 2007). data: shape (T, d)."""
T, d = data.shape
nu0 = nu0 or float(d + 2)
Psi0 = Psi0 if Psi0 is not None else np.eye(d) * 1e-4
R = np.zeros((T + 1, T + 1))
R[0, 0] = 1.0
model = MultivariateStudentT(d, kappa0=kappa0, nu0=nu0, Psi0=Psi0)
h = 1.0 / hazard_rate
maxes = np.zeros(T)
p_short = np.zeros(T)
for t in range(T):
x = data[t]
pred = np.exp(model.log_pred_prob(x))
R[1:t+2, t+1] = R[:t+1, t] * pred * (1 - h)
R[0, t+1] = np.sum(R[:t+1, t] * pred * h)
ev = R[:t+2, t+1].sum()
if ev > 0:
R[:t+2, t+1] /= ev
model.update(x)
maxes[t] = np.argmax(R[:t+2, t+1])
p_short[t] = R[:min(10, t+2), t+1].sum()
changepoints = [t for t in range(1, T) if maxes[t] < 5 and maxes[t-1] > 20]
return R, changepoints, maxes, p_short
Run it with a weakly informative prior. "Weakly informative" here has a specific meaning: the structure is uninformative (diagonal Psi0 assumes zero cross-asset correlation, letting the data determine all covariance structure), but the scale is calibrated to each asset's historical return variance rather than set to an isotropic identity matrix. A pure identity prior would be miscalibrated: oil returns are an order of magnitude more volatile than the USD index, so treating them as equally scaled biases the predictive distribution from the start. The diagonal-but-variance-scaled choice is the least informative prior that is still numerically sensible:
signal = log_ret.values # (T, 5)
d = signal.shape[1]
# Weakly informative diagonal prior scaled to typical return variance
nu0 = float(d + 2)
Psi0 = np.diag(log_ret.std().values**2 * (nu0 - d - 1))
R, changepoints, maxes, p_short = bocd_multivariate(
signal, hazard_rate=500, nu0=nu0, Psi0=Psi0
)
print(f"Found {len(changepoints)} changepoints")

Running this on 5,617 aligned trading days (October 2003 to April 2026) finds 69 changepoints. The chart above shows the 15 where the preceding regime lasted at least 125 trading days, the durable macro transitions rather than short-lived dislocations.
What Just Happened?

The GIF above shows the model processing Liberation Day (April 2, 2025) in real time. As S&P 500 and oil prices move together on the left, , the probability that the current regime started within the last 10 days, rises on the right. No single asset produced a move large enough to reset the run-length distribution on its own. Jointly, the five assets painted an unambiguous picture, and a changepoint was registered on April 3, one day after the tariff announcement.
Validating the 69 Detections
We looked up all 69 detected changepoints against documented market and macro events, then applied a stricter test: does the event actually explain a shift across multiple asset classes simultaneously, or did we just find the nearest headline on that date? (Full detection log with event dates and lags.)
| Classification | Count | Share |
|---|---|---|
| CLEAR: named macro event with cross-asset impact, within ±5 trading days | 47 | 68.1% |
| PLAUSIBLE: genuine regime-adjacent period, no exact-date catalyst | 10 | 14.5% |
| TENUOUS: event found by calendar search; limited cross-asset evidence | 12 | 17.4% |
| UNCLEAR: no identifiable driver | 0 | 0% |
The TENUOUS cases are worth owning rather than papering over. They fall into three patterns: routine Fed hikes in a well-telegraphed cycle (2005-11-01, 2005-12-14) where markets had priced the move months in advance; single-asset spikes that did not propagate (Refco fraud in 2005, the gold crash in 2013, the Abqaiq oil spike in 2019, the Nasdaq mini-crash in 2020); and detections that are effectively redundant: the model firing again days into a crisis that had already been detected (Ukraine commodity shock five days after the invasion; China Politburo two days after the PBOC bazooka). The remaining two, the Rosengren hawkish speech (2016) and the US Tax Cuts passage (2017), caused equity sector rotations but not the kind of durable cross-asset regime break these five series would be expected to show.
The table below shows the 12 detections that pass both the ≥125-day regime filter and the CLEAR or PLAUSIBLE classification, the durable macro transitions visible in the chart above.
| Date | Class | Event |
|---|---|---|
| 2007-02-27 | CLEAR | "Shanghai Sneeze": CSI 300 crashed 9%; Dow's worst day since 9/11 |
| 2008-07-30 | CLEAR | HERA signed: Treasury backstop for Fannie/Freddie; IndyMac had just failed |
| 2009-12-04 | PLAUS | Surprise November jobs report: dollar surged, gold fell; no single regime catalyst |
| 2011-01-28 | CLEAR | Egypt "Friday of Anger": Tahrir Square; regional contagion spiked oil |
| 2012-06-01 | CLEAR | NFP +69k miss + Spain/Greece crisis: Dow -274 pts, worst day of 2012 |
| 2014-01-02 | PLAUS | EM currency crisis onset: Argentina, Turkey, Brazil selloff; worst EM month since 2008 |
| 2014-09-04 | CLEAR | ECB surprise cut to 0.05% + ABS purchases: euro fell 1%+; EU stocks rallied |
| 2015-08-24 | CLEAR | China "Black Monday": Shanghai -8.49%; Dow opened -1,000 pts; oil at 6-year low |
| 2019-06-20 | CLEAR | Fed dovish pivot: Powell signalled rate cuts; bonds surged, dollar fell |
| 2023-10-13 | CLEAR | Hot CPI + intraday reversal; within one week of Hamas October 7 attack |
| 2024-08-02 | CLEAR | Weak jobs (114k vs 175k expected) + BOJ rate hike triggers yen carry trade unwind |
| 2025-10-10 | CLEAR | Trump 100% tariff on Chinese imports + China rare earth export controls; S&P -2.7% |
The 47 CLEAR detections span every major crisis type: monetary policy pivots (Fed, ECB, PBOC), geopolitical shocks (Ukraine, Libya, Iran), financial crises (2008 cluster, Volmageddon, Omicron), and commodity regime breaks (OPEC 2014, silver crash 2011, oil spikes). The 10 PLAUSIBLE cases are periods where a genuine cross-asset dislocation was underway but no single catalyst fell on the exact day: the model was detecting something real without a clean label for it.
The Iran War: Where Univariate Fails

The chart above shows around the Iran war event (February 28, 2026) for both models.
The multivariate model's reached 100% on March 2, 2026, two days after the war began. The univariate model's never left its 0.41% baseline across the entire window. The teal line spikes to certainty; the amber line does not move.
The mechanism is straightforward. When Iran was attacked on February 28, oil futures spiked sharply (supply disruption fears), gold rose as a safe haven, emerging market equities fell on regional risk-off, and the dollar strengthened. The S&P 500 also moved, but was already in a high-volatility regime from Liberation Day. Individually, none of these moves was dramatic enough to reset a run-length distribution conditioned on high volatility.
Jointly, however, the cross-asset co-movement pattern was distinctive. Within the tariff-volatility regime, oil and EM equities had not been moving in opposite directions; their covariance was different. The Iran war produced a covariance signature that Psi (which accumulates the regime's observed covariance) had not seen. The predictive likelihood dropped sharply, and the run-length distribution shifted weight towards .
Regime Covariance: What Changes Between Eras

The Psi matrix at the end of a regime contains a scatter matrix reflecting the cross-asset covariance seen during that period. Comparing Psi / (nu - d - 1) across regimes reveals how asset relationships change.
The two longest detected regimes were October 2003 to October 2005 (482 days) and November 2016 to November 2017 (264 days). Their correlation structures differ in revealing ways.
In the early regime, S&P 500 and EM equities were tightly correlated (0.70): the post-Iraq-war global reflation trade lifted both together. Gold and the dollar had a strong negative relationship (-0.68), consistent with the textbook safe-haven dynamic. Oil moved on supply factors rather than macro risk.
In the 2016–2017 regime (Trump's first year), the S&P 500–EM correlation fell to 0.56 as diverging growth expectations separated the two. Gold turned negatively correlated with the S&P 500 (-0.23): markets began treating gold as a hedge against equity risk rather than purely a dollar hedge. The gold–dollar relationship remained negative (-0.47) but weakened. These structural differences are exactly what Psi accumulates, and what makes a new observation feel "out of regime" when the covariance pattern shifts.
Run-Length Posterior


The heatmap shows the full run-length posterior over time. Each column is the probability distribution over regime ages at a given date. Bright horizontal bands indicate extended periods where the run-length posterior sits at a stable value: the model is confident the regime started at a particular point in the past. The bright streaks that reset to the bottom are the detected changepoints: the posterior collapses and the run-length distribution starts fresh.
Going Deeper
What Changed (and What Did Not)
The outer bocd_multivariate function is nearly character-for-character identical to the bocd function from our previous post. The three-step loop (predict, update R, normalise) is unchanged. The hazard function is unchanged. The run-length matrix R has the same shape and semantics.
What changed is entirely inside the predictive model. In Post 45, each run-length hypothesis carried four scalar parameters: mu, kappa, alpha, beta. Here, each hypothesis carries a mean vector mu of shape (d,), a precision parameter kappa, a degrees-of-freedom parameter nu, and a d × d scale matrix Psi.
Psi is the key addition. It is a scatter matrix that grows with each observation within the current regime. The rank-1 update Psi += (kappa / (kappa + 1)) * outer(x - mu, x - mu) accumulates the cross-asset covariance structure of returns. When a new regime starts (run length resets to 0), Psi resets to the prior Psi0.
The Cholesky Trick
In the univariate case, computing required evaluating a scalar Student-t pdf, which is . In the multivariate case, we need two things for each run-length hypothesis:
- : the log-determinant of the scale matrix.
- : the quadratic form.
Done naively, each requires work per hypothesis. By computing the Cholesky decomposition , we get the log-determinant as and the quadratic form via forward substitution. NumPy's np.linalg.cholesky and np.linalg.solve both accept batched 3D inputs, so we decompose all n_rl scale matrices in a single vectorised call rather than looping.
Prior Selection
The choice nu0 = d + 2 is the minimum value that gives a proper inverse-Wishart distribution with a finite mean. The inverse-Wishart requires for existence and for a finite mean. Setting sits just above this boundary: the data will dominate almost immediately.
For the scale matrix, the choice Psi0 = diag(sigma^2) * (nu0 - d - 1) is motivated by the prior mean of the inverse-Wishart distribution:
Setting Psi0 = diag(sigma^2) * (nu0 - d - 1) makes this prior mean equal to diag(sigma^2), a diagonal covariance equal to the observed marginal variances, with no prior assumption about cross-asset correlation. Off-diagonal entries of Psi0 are zero; the posterior learns the actual correlation structure from the data within each regime.
Hazard Rate
The univariate BOCD found 7 changepoints on the same data window (2003–2026). The multivariate model finds 69. This tenfold difference reflects a genuine feature of higher-dimensional observations: the joint predictive likelihood is a product over all five series, so it drops faster when the data deviates from the current regime in multiple dimensions simultaneously.
The key implication from the validation above is that raising hazard_rate is not the right response to the higher count. Of the 69 detections, 57 are CLEAR or PLAUSIBLE, genuinely real. The 12 TENUOUS cases are not noise in the random-variation sense; they are detections where the model responded to a real cross-asset co-movement that happened to coincide with a minor or redundant event. The 2022 cluster alone yields five legitimate detections: Russia invading Ukraine, the first Fed rate hike, the 50bp hike, the 75bp hike, and the UK gilt crisis. (The commodity shock on 2022-03-01, five trading days after the invasion, is the one 2022 detection we classify as TENUOUS: the invasion had already been detected and the commodity ripple is the same regime continuing.)
The right tool for managing chart density is the visualization filter, not hazard_rate. The chart in the Quick Win section shows only changepoints where the preceding regime lasted at least 125 trading days, which selects for durable macro transitions and filters out short-lived dislocations. For real-time monitoring, run the model at hazard_rate=250 or hazard_rate=500 and monitor directly. For a retrospective chart of major regime breaks, apply the minimum-regime-length filter instead.
Memory and Performance
At each time step , we store a Psi tensor of shape . This grows quadratically in time: at observations with , peak storage is bytes, approximately 1.1 MB.
Runtime scales as in the worst case. In practice, the batched NumPy calls are efficient: on a modern laptop, 5500 observations with runs in under 5 minutes.
For streaming applications, consider pruning run-length hypotheses below a probability threshold. Dropping hypotheses where R[r, t] < 1e-10 keeps the active hypothesis count bounded without meaningfully affecting the posterior, reducing memory and runtime to effectively .
The Math
NIW Sufficient Statistics Update
After observing data points within a regime, the NIW posterior hyperparameters are:
The update method in MultivariateStudentT applies a sequential version of these updates: each new observation extends each hypothesis by one, then prepends the prior as the new run_length = 0 hypothesis.
Predictive Distribution
The marginal predictive distribution for a new observation given observations is a multivariate Student-t:
This is the exact multivariate analogue of the univariate Student-t from Post 45. The degrees of freedom shrink as dimensionality increases, making the tails heavier in higher dimensions. For , this reduces to the formula in the StudentT class from our previous post.
The log pdf of this distribution is:
where and . This is exactly what log_pred_prob computes, split across the variables df, sf, S, log_det, and quad.
Why the NIW is Conjugate
The Normal-Inverse-Wishart is conjugate to the multivariate Gaussian with unknown mean and covariance. This means the posterior after observing is also NIW, with updated hyperparameters as above. Conjugacy is what lets us compute the predictive distribution in closed form at each step, without numerical integration. The multivariate Student-t predictive distribution drops out analytically by marginalising the unknown over the NIW posterior.
Where This Comes From
Adams and MacKay's 2007 paper is the foundation here, and we covered it in detail in Post 45. One of the paper's central design decisions was making the algorithm modular: the underlying predictive model is swapped out independently from the changepoint logic. The NIW extension follows directly from this design.
The Normal-Inverse-Wishart prior itself was not introduced in the BOCD paper. It is a classical result in Bayesian multivariate statistics, covered in Murphy (2012) Chapter 4.6 as the natural conjugate for the multivariate Gaussian. Turner, Saatci, and Rasmussen (2009) were among the first to use NIW within the BOCD framework, noting that it enables exact closed-form inference in the multivariate case, which particle filtering approaches from Fearnhead and Liu (2007) had approximated.
The multivariate Student-t predictive distribution we use is discussed in Bishop (2006) PRML Chapter 2.3.7 as the marginal distribution obtained by integrating out the unknown Gaussian parameters under a conjugate prior.
References
- Adams & MacKay (2007): Bayesian Online Changepoint Detection
- Murphy (2012): "Machine Learning: A Probabilistic Perspective", Chapter 4.6 (Normal-Inverse-Wishart distribution)
- Bishop (2006): "Pattern Recognition and Machine Learning", Chapter 2.3.7 (multivariate Student-t as predictive distribution)
- Turner, Saatci & Rasmussen (2009): Gaussian Process Change Point Models
Appendix Full Detection Log
Every detected changepoint with the actual event date and detection lag. Lag = BOCD detection date minus event date: negative means the model fired before the headline; positive means it lagged. PLAUSIBLE cases have no single event date: the model detected a real cross-asset dislocation with no single identifiable catalyst.
Of the 47 CLEAR detections: 30 fired on the same day, 8 fired 1–2 days late (the next session processed the full co-movement), and 9 fired 1–3 days early (cross-asset tension was already building before the headline). The 10 PLAUSIBLE cases are regime-adjacent periods with no single-date trigger. The 12 TENUOUS cases are detections where a cross-asset move occurred but the nearest calendar event does not constitute a genuine macro regime shift for all five series: single-asset spikes, routine hikes in a fully telegraphed cycle, or redundant detections within the same ongoing crisis.
View all 69 changepoints with event dates and detection lags
| BOCD Detection | Event Date | Lag | Class | Event |
|---|---|---|---|---|
| 2005-10-07 | 2005-10-10 | −3 days (early) | TENUOUS | Refco fraud disclosure: single broker's stock collapsed; S&P, oil, gold, EM, USD not materially moved |
| 2005-11-01 | 2005-11-01 | same day | TENUOUS | Fed FOMC hike +25bp: 8th consecutive hike in a fully telegraphed cycle; markets had priced it months ahead |
| 2005-12-14 | 2005-12-13 | +1 day | TENUOUS | Fed FOMC hike +25bp: routine hike in same telegraphed cycle; no new macro information |
| 2006-06-13 | 2006-06-13 | same day | CLEAR | May–June 2006 EM/commodity correction: inflation fears; Saudi stocks −50% |
| 2007-02-27 | 2007-02-27 | same day | CLEAR | "Shanghai Sneeze": CSI −9%; Dow worst day since 9/11 (−416 pts) |
| 2007-06-07 | 2007-06-07 | same day | CLEAR | Bond rout (10Y UST 5.12%) + Bear Stearns suspended hedge fund withdrawals |
| 2007-08-06 | 2007-08-06 | same day | CLEAR | Quant crisis / credit crunch: Bear Stearns bankruptcies; ECB/Fed emergency liquidity |
| 2007-11-12 | n/a | n/a | PLAUS | Mid-credit-crisis deepening; oil near $100; no single catalyst on exact date |
| 2008-07-30 | 2008-07-30 | same day | CLEAR | HERA signed: Treasury backstop for Fannie/Freddie; IndyMac had just failed |
| 2008-09-15 | 2008-09-15 | same day | CLEAR | Lehman Brothers bankruptcy: largest US bankruptcy ever |
| 2008-09-17 | 2008-09-17 | same day | CLEAR | Reserve Primary Fund "broke the buck"; $144.5B withdrawn from money markets in one day |
| 2009-12-04 | n/a | n/a | PLAUS | Surprise Nov jobs report (+11k vs −100k expected); dollar surged, gold fell sharply |
| 2010-05-10 | 2010-05-10 | same day | CLEAR | EU/ECB/IMF €750B Greek bailout announced; Flash Crash (May 6) aftermath |
| 2010-06-21 | 2010-06-21 | same day | CLEAR | China announced end of dollar peg (Jun 19–21): RMB managed float resumed |
| 2010-07-01 | n/a | n/a | PLAUS | Dodd-Frank passage nearing (signed Jul 21); financial-sector regime shift underway |
| 2011-01-28 | 2011-01-28 | same day | CLEAR | Egypt "Friday of Anger": Tahrir Square; regional contagion spiked oil |
| 2011-02-22 | 2011-02-22 | same day | PLAUS | Libya escalation: primarily an oil spike; S&P, gold, EM, USD effect limited and short-lived |
| 2011-05-05 | 2011-05-05 | same day | CLEAR | Commodity crash: silver −10% (−27% from peak); oil <$100; CME margin hikes |
| 2011-08-08 | 2011-08-08 | same day | CLEAR | "Black Monday 2011": S&P US debt downgrade; Dow −634 pts (−5.55%) |
| 2011-09-23 | 2011-09-21 | +2 days | CLEAR | Operation Twist announced; S&P −3.19%; gold worst monthly decline |
| 2012-06-01 | 2012-06-01 | same day | CLEAR | Worst day of 2012: NFP +69k miss; Spain/Greece crisis; Dow −274 pts |
| 2012-07-03 | 2012-07-05 | −2 days (early) | CLEAR | ECB cut to 0.75% (Jul 5); Draghi "whatever it takes" buildup |
| 2012-12-28 | 2012-12-28 | same day | PLAUS | US fiscal cliff climax: primarily a US equity event; oil, gold, EM, USD did not show a lasting break |
| e | 2013-04-15 | 2013-04-15 | same day | TENUOUS |
| 2013-06-21 | 2013-06-19 | +2 days | CLEAR | Taper Tantrum: Bernanke signalled tapering (Jun 19); 10Y +40bp; EM currencies collapsed |
| 2014-01-02 | n/a | n/a | PLAUS | EM stress beginning: worst EM selloff in 5 years (Argentina, Turkey, Brazil) |
| 2014-09-04 | 2014-09-04 | same day | CLEAR | ECB surprise cut to 0.05% + ABS purchases; euro −1%+; EU stocks rallied |
| 2014-10-09 | 2014-10-09 | same day | CLEAR | Global growth scare: Dow worst day of 2014; Ebola fears; 10Y <2.3% |
| 2014-11-21 | 2014-11-21 | same day | CLEAR | PBOC surprise rate cut (first since 2012): 40bp; global equities rallied sharply |
| 2014-11-28 | 2014-11-27 | +1 day | CLEAR | OPEC Vienna "no cut": WTI −7% to <$70; oil bear market confirmed |
| 2015-02-03 | 2015-02-04 | −1 day (early) | CLEAR | ECB removed Greek bond collateral waiver (Feb 4); Greek banks forced onto ELA |
| 2015-08-24 | 2015-08-24 | same day | CLEAR | China "Black Monday": Shanghai −8.49%; Dow opened −1,000 pts; oil 6-year low |
| 2016-02-11 | 2016-02-11 | same day | CLEAR | Global market low: oil multi-year low; gold +4.4%; Yellen mentioned negative rates |
| 2016-06-24 | 2016-06-24 | same day | CLEAR | Brexit result: GBP −10% (31-year low); Dow futures −900 pts; $2T+ wiped globally |
| 2016-09-09 | 2016-09-09 | same day | TENUOUS | Rosengren hawkish speech: single Fed official; 2% equity drop fully reversed within days; not a macro regime |
| 2016-11-09 | 2016-11-09 | same day | CLEAR | Trump election result: Dow futures −900 pts then +257 close; USD surged; MXN −11% |
| 2017-11-30 | 2017-12-01 | −1 day (early) | TENUOUS | US Tax Cuts Senate passage: equity sector rotation within US equities; oil, gold, EM, USD not in a new regime |
| 2018-02-02 | 2018-02-02 | same day | CLEAR | "Volmageddon" trigger: wage growth surprise; VIX 17→50 on Feb 5; XIV destroyed |
| 2018-06-14 | 2018-06-13 | +1 day | CLEAR | Fed hike to 1.75–2.0%; EM currencies under pressure; Turkish lira/ARS in crisis |
| 2018-10-10 | 2018-10-10 | same day | CLEAR | Tech rout: Dow −831 pts; 10Y at 7-year high (3.23%); Italy budget fears |
| 2018-11-13 | 2018-11-13 | same day | TENUOUS | Oil −6% + Apple −4.8%: same Q4 2018 regime already detected in October; continuation, not a new break |
| 2019-06-20 | 2019-06-19 | +1 day | CLEAR | Fed dovish pivot: Powell signalled cuts (Jun 19 FOMC); bonds surged, dollar fell |
| 2019-08-05 | 2019-08-05 | same day | CLEAR | China yuan breaks 7/dollar; Dow −767 pts (worst of 2019); trade war escalation |
| 2019-09-16 | 2019-09-14 | +2 days | TENUOUS | Abqaiq drone attack: Brent +15–20% intraday then largely reversed within a week; oil spike, not a macro regime shift |
| 2020-02-25 | 2020-02-25 | same day | CLEAR | COVID-19 global panic onset: 10Y to record 1.31%; oil 13-month low |
| 2020-08-11 | n/a | n/a | PLAUS | Sputnik V approval + tech/value rotation beginning; cross-asset regime shift |
| 2020-09-03 | 2020-09-03 | same day | TENUOUS | Nasdaq mini-crash: single worst equity day of the summer; no lasting cross-asset regime break |
| 2021-07-27 | 2021-07-27 | same day | TENUOUS | China tech crackdown: EM equities affected but S&P, oil, gold, USD not meaningfully repriced; EM-local story |
| 2021-11-11 | 2021-11-10 | +1 day | CLEAR | US CPI 6.2% (Nov 10): 31-year high; 10Y surged; rate hike expectations repriced |
| 2021-11-26 | 2021-11-26 | same day | CLEAR | Omicron variant: Dow −900 pts (worst day of year); oil −13%; airlines −9% |
| 2022-02-24 | 2022-02-24 | same day | CLEAR | Russia invades Ukraine: oil >$100; gold >$1,900; Moscow MOEX −33% |
| 2022-03-01 | 2022-03-01 | same day | TENUOUS | Ukraine commodity ripple: 5 trading days after the invasion was already detected (2022-02-24); same regime continuing |
| 2022-03-16 | 2022-03-16 | same day | CLEAR | First Fed rate hike of 2022 (+25bp); signalled 7 hikes in 2022 |
| 2022-05-02 | 2022-05-04 | −2 days (early) | CLEAR | Fed +50bp (May 4): largest hike in 22 years; Powell ruled out 75bp (later reversed) |
| 2022-06-16 | 2022-06-15 | +1 day | CLEAR | Fed +75bp (Jun 15): largest since 1994; response to 8.6% CPI |
| 2022-10-24 | 2022-10-20 | +4 days | CLEAR | Liz Truss resignation (Oct 20) + gilt crisis resolution; BoE bought £100B+ in gilts |
| 2023-10-13 | 2023-10-13 | same day | CLEAR | Hot CPI + dramatic intraday reversal +2.6%; within one week of Hamas Oct 7 attack |
| 2023-10-18 | n/a | n/a | PLAUS | 10Y approaching 5% (hit Oct 19); Israel-Gaza ongoing; bond regime shift |
| 2024-08-02 | 2024-08-02 | same day | CLEAR | Weak jobs (114k vs 175k expected) + BOJ rate hike triggers yen carry trade unwind |
| 2024-09-24 | 2024-09-24 | same day | CLEAR | China PBOC "bazooka": rate cuts + ¥500B equity facility; CSI 300 +16% in one week |
| 2024-09-26 | 2024-09-26 | same day | TENUOUS | China Politburo follow-up: 2 trading days after PBOC bazooka already detected (2024-09-24); same stimulus regime |
| 2024-10-31 | n/a | n/a | PLAUS | October month-end drift; Middle East tensions + AI cost concerns; no sharp catalyst |
| 2024-12-18 | 2024-12-18 | same day | CLEAR | "Hawkish cut": Fed cut 25bp but dot plot reduced 2025 cuts to 2; Dow fell 10th day |
| 2025-03-06 | n/a | n/a | PLAUS | Feb jobs above consensus (228k vs 140k); tariff uncertainty mounting |
| 2025-04-03 | 2025-04-02 | +1 day | CLEAR | Liberation Day: Trump ~20% universal tariffs; S&P −5%; $3T+ global market cap destroyed |
| 2025-10-10 | 2025-10-10 | same day | CLEAR | Trump 100% tariff on Chinese imports + China rare earth export controls; S&P −2.7% |
| 2026-01-02 | 2026-01-02 | same day | CLEAR | Bond vigilante "yield shock": 30Y UST multi-month high; fiscal deficit fears; global selloff |
| 2026-01-30 | 2026-01-30 | same day | CLEAR | Fed holds + Trump nominates Warsh (hawkish) as next Fed Chair; shutdown deadline same day |
| 2026-03-02 | 2026-02-28 | +2 days | CLEAR | US-Israel strikes on Iran (Feb 28): Brent +40% to $106; NYSE −6%; Nikkei −11% |
Interactive Tools
- Normal Distribution Calculator: The univariate Gaussian that the multivariate observation model in this post generalises
- Gaussian Process Playground: Another Bayesian online method using GP priors
- Kelly Criterion Calculator: size positions per detected regime
- Black-Scholes Calculator: reprice options as volatility regimes shift
- Drawdown Calculator: quantify drawdowns around the detected crises
Related Posts
For time-varying financial volatility, the GARCH Python tutorial compares fitted conditional variance and forecasts.
The Kelly criterion Python tutorial examines how estimation uncertainty and drawdowns affect position sizing.
Changepoint Detection: Finding Regime Shifts in Financial Data: The first post in this series: offline PELT on S&P 500 data
Bayesian Online Changepoint Detection in Python: The second post: univariate BOCD that this post extends
Hidden Markov Models: When Clusters Have Memory: Probabilistic regime detection with smooth transitions rather than hard boundaries
MCMC for Mixture Models: Inferring Earthquake Regimes: Bayesian regime detection in count data
Cointegration and Pairs Trading: Uses the same 5 asset classes in a different quant-finance framework
Modern Portfolio Theory from Scratch: The Efficient Frontier: Detecting the regime shifts that break a static efficient frontier
LLM Use Cases for Asset Managers: Where cross-asset regime detection meets the generative-AI workflow asset managers are actually asking for: desk-level macro narratives, mandate-drift checks, and multi-manager oversight.
Time Series as Dynamical Systems: Delay Embedding: Regime shifts across asset classes are transitions in the very state space delay embedding reconstructs.









