Prevent look-ahead bias
I applied a one-day prior-close signal lag before assigning each regime.
Statistical Analytics · Data Validation
Macro-Financial Datasets
A regime analysis that tests bias, threshold sensitivity, uncertainty, and cross-system agreement.
Role
Data Analyst
Tools
Python · PostgreSQL · Tableau
Sample
1,111 trading days
Stability
6 threshold combinations
Control
5,000 resamples
My approach
Challenge the classification
I tested six VIX and yield threshold combinations and compared asset metrics by regime.
Make uncertainty visible
I used 5,000 moving-block bootstrap resamples and reconciled outputs across Python and PostgreSQL.
Analytical trust
A result is useful only when the method is stable.
The validation path challenges the finding before it is presented.
01Time-series data
↓
02One-day lag
↓
03Regime classification
↓
04Sensitivity tests
↓
05Bootstrap uncertainty
↓
06System reconciliation
GLD led in Tightening.
The initial result identifies a regime relationship.
Does one threshold create it?
Six threshold combinations test the classification.
The pattern remained consistent.
The stability test supports the descriptive finding.
Decision evidence
Which findings survive challenge?
Point estimates are shown with method controls and uncertainty, not presented as forecasts.
01 · Regime comparison
How did assets behave by regime?
Why this visual exists. Return, risk, drawdown, Sharpe ratio, and correlation are read together across four lagged regimes.
Finding
GLD has the highest Sharpe ratio at 1.95 in Elevated risk. SPY has the deepest maximum drawdown at −13.3% in the same regime.
Interpretation
No single metric defines performance. Return must be read with volatility, drawdown, and cross-asset behaviour.
Decision
Use the dashboard to select claims for robustness checks, not to issue an investment recommendation.
02 · Point estimate
Which regime differences are largest?
Why this visual exists. A common return view makes direction and magnitude easy to compare before testing uncertainty.
Finding
SPY is 66.8% annualised in Stress, while IEF and GLD are −20.9% and −22.1%. GLD is 31.5% in Elevated risk.
Interpretation
These are descriptive annualised sample means. The Stress estimate can be unstable because regime samples are limited.
Decision
Do not use magnitude alone. Check sensitivity and sampling uncertainty before retaining the claim.
03 · Uncertainty
How precise are the estimates?
Why this visual exists. Moving-block bootstrap intervals preserve short time dependence and expose imprecise regime estimates.
Finding
Several intervals cross zero, and Stress intervals are especially wide across assets.
Interpretation
The point estimate can look strong while uncertainty remains too large for a confident directional claim.
Decision
Report ranges and evidence limits. Treat the analysis as descriptive, not predictive.
Source: original project dashboard and figures. Intervals use 5,000 deterministic moving-block resamples and are not forecast ranges.
Related concepts
Explore how the result is challenged.
Method
Analytical rigour
Explore concept →Trust
Data trust
Explore concept →Metrics
Measurement design
Explore concept →Lag the signal before classifying the return.
The regime uses information available at the previous close. This prevents the current return from using a same-day closing signal.
View implementation on GitHub ↗
result["dgs10_change_63d_pp"] = (
result["dgs10"] - result["dgs10"].shift(YIELD_LOOKBACK_DAYS)
)
result["signal_vix"] = result["vix"].shift(1)
result["signal_dgs10"] = result["dgs10"].shift(1)
result["signal_dgs10_change_63d_pp"] = (
result["dgs10_change_63d_pp"].shift(1)
)