RAHUL KHANKE / UNIVERSITY OF GLASGOW / MSc DATA SCIENCE · 2025
Where do markets change their minds?
Bayesian change-point analysis across national stock indices and a five-crypto research cohort. Explore synchronized weekly returns, compare detected structural shifts and test whether findings survive changes in the model's assumptions.
9 NATIONAL INDICES5 SELECTED CRYPTOASSETSREAL HISTORICAL PRICESMODEL SENSITIVITY
02 / COMPARATIVE EVIDENCE
The Regime Atlas
Select markets, then run analysis to fetch source-attributed history.
MARKETS ANALYZED—
WEEKLY OBSERVATIONS—
MODEL SETTINGS9
A / WEEKLY LOG RETURNS (%, NOT PRICE LEVELS)
Your selected markets will appear here.
B / SENSITIVITY & CONSENSUS
Changes are judged against nine hazard/noise combinations; they are hypotheses, not independently confirmed events.
C / CROSS-MARKET SYNCHRONIZATION
Common week-end dates make it possible to inspect temporal overlap without forward-filling non-trading days.
Research, not investment advice. Yahoo Finance is an unofficial source with variable coverage and downtime; market data timestamps, calendar rules, model parameters and any missing series are reported. Reliability here means parameter stability, not forecast accuracy or out-of-sample validation.
03 / THE RESEARCH QUESTION
Does the pattern travel?
A single well-tuned chart can be persuasive for the wrong reasons. This study asks whether candidate structural breaks are robust across national equity markets, cryptocurrency returns and alternate model assumptions. The sensitivity panel repeats the same Bayesian recursion under three hazard settings and three observation-variance scales.
Equity trading days and 24/7 crypto markets cannot be aligned one-to-one. Using each market's last observed close within a common Sunday–Saturday UTC window avoids inventing missing daily returns.
04 / METHOD & CAUTION
A retrospective model, not a crystal ball.
We compute weekly percentage log returns, then apply a Gaussian mean-shift Bayesian online change-point recursion. The fixed observation variance is estimated from each full sample, so these analyses are retrospective and must not be described as strictly ex-ante forecasting.
Consensus measures how frequently nearby peaks appear among nine model configurations; it is not a statistically calibrated confidence probability or an accuracy score. For a dissertation-grade result, add reference event labels, alternative change-point models, sample splits and archived input data.