Replications on the estate
Research
Each note re-runs a published result in financial economics on the data in this repository: the published specification first, then at most one pre-declared variant. The note states the paper's claim as the paper's, states exactly how this sample differs, and reports the result as it comes out. Every number is computed at build time from the estate; nothing is typed in. A replication that fails is reported as a headline, not buried as a robustness footnote.
Schularick and Taylor (2012), American Economic Review
Reproduces. On the current JST vintage, with twelve more years of data than the paper had, lagged real loan growth still predicts banking-crisis onsets: the lag-two coefficient is large and significant, the five-lag block is jointly significant, and the credit model beats the money model on fit and discrimination in both samples.
Baron, Verner and Xiong (2021), Quarterly Journal of Economics
Reproduces, on the authors' own replication panel, and the note says so rather than claiming an out-of-sample test. Bank-equity crashes without panics are followed by real output losses; crashes with panics by substantially larger ones.
Shiller (1981), American Economic Review
The direction survives, the magnitude does not. On the data through the present, price volatility still exceeds the bound implied by realized dividends, but by far less than the factor of five to thirteen the paper reported, and the note shows how much of that depends on the terminal condition.
Reinhart and Rogoff (2010) vs Herndon, Ash and Pollin (2014)
Herndon, Ash and Pollin reproduce; Reinhart and Rogoff do not. On the JST panel over the papers' own window, growth above 90% of GDP is positive under both weighting schemes, growth declines with debt across the whole range, and there is no cliff at 90%.
Cruces and Trebesch (2013), AEJ: Macroeconomics
The headline moment reproduces almost exactly on the original cases. The paper's central price-of-haircuts regressions are untestable on this estate, because it carries no market-reaccess dates and no sovereign spread panel, and the note confines itself to what the file can support.
Borio and Lowe (2002), BIS Working Papers
Half reproduces. On the BIS's own published credit-to-GDP gap scored against the Laeven-Valencia chronology, the gap does lead banking crises, but the low noise-to-signal ratio the paper reported does not reproduce at its own thresholds.
Fama and French (2015, 2016), Journal of Financial Economics and Review of Financial Studies
Reproduces. Most industry alphas are individually small, the significant ones are mostly negative, and yet the joint test still rejects the model on the balanced panel: industries remain the hard test the authors said they were.
Drehmann and Juselius (2014), International Journal of Forecasting
Does not reproduce, and that is the headline. On the horizons and crises this panel can test, all from one crisis wave, the credit-to-GDP gap beats the debt service ratio at every horizon; only the paper's short-fuse shape, the DSR improving as the crisis nears, survives.
The scripts that compute every figure live in scripts/build/build_research_*.py in the repository, and each note documents its method and limitations inline.