Replication data for: Loss Aversion and the Asymmetric Exchange-Rate Sensitivity of Money Demand: Bayesian Evidence from Turkey
Description
This dataset accompanies the manuscript “Loss Aversion and the Asymmetric Exchange-Rate Sensitivity of Money Demand: Bayesian Evidence from Turkey,” submitted to the Journal of International Money and Finance. It provides the data, code, MCMC chains, and intermediate outputs required to reproduce the empirical analysis in full. The archive includes the model-ready monthly Turkish panel for 2003-01 through 2025-12, comprising log real M2, the calendar-adjusted Industrial Production Index, the spliced policy rate, the TL/USD exchange rate, the GJR-EGARCH(1,1)-Student-t conditional-variance series, and the TP.KM dollarization ratio. The original raw pulls from the CBRT- Electronic Data Delivery System (EDDS) and Turkstat are retained, together with the splice conventions used to bridge institutional revisions in the policy-rate and monetary-aggregate series. The PyMC v5 model code covers the main Bayesian specification, the prior-sensitivity comparison (S1/S2/S3), the hierarchical regime-dependent extension, and the dollarization-interaction robustness specification. MCMC trace files in NetCDF format (ArviZ-compatible) and tabulated posterior summaries are provided for every specification. The full ARDL/NARDL benchmark code is included, with AIC lag selection, Pesaran–Shin–Smith bound testing, error-correction representation, and the WUI/GEPU/VIX uncertainty-augmentation analysis. The matplotlib code used to produce the six figures in the manuscript is also bundled. The empirical analysis recovers the prospect-theory loss-aversion coefficient λ from emerging-market money-demand data: the pooled posterior mean is 2.17 with 94% credible interval [1.18, 3.64], statistically indistinguishable from the Tversky–Kahneman 1992 laboratory benchmark of 2.25. The diminishing-sensitivity coefficient α = 0.83 rejects the linear-asymmetric restriction implicit in the nonlinear ARDL literature (posterior probability of α < 1 is 100%). A hierarchical regime-conditional extension locates the behavioral channel in the post-2018 stressed exchange-rate-trend regime. Reproduction requires Python 3.11 or later; the package requirements and pipeline run order are specified in the README.md included in the archive. The main MCMC estimation completes in approximately 8–12 minutes on a modern laptop. All raw macroeconomic series are publicly available from the TCMB EVDS and TÜİK; no proprietary or sensitive data is included. The series codes used and the splice conventions are documented in data/data_pull_checklist.md.
Files
Steps to reproduce
1. Install Python 3.11 or later and the packages listed in requirements.txt (pymc≥5.10, arviz≥0.16, arch≥6.2, statsmodels≥0.14, evds≥2.0, pandas, numpy, scipy, matplotlib). 2. To refresh the raw TCMB pulls, set the EVDS_API_KEY environment variable to a valid EVDS3 key (registration is free at https://evds2.tcmb.gov.tr/) and run python code/data_prep/pull_evds.py. To reproduce from the supplied raw CSVs, skip this step. 3. Run python code/data_prep/build_panel.py to assemble the model-ready monthly panel. The output is written to data/processed/panel_monthly.csv. 4. Run python code/data_prep/garch_volatility.py to estimate the GJR-EGARCH(1,1)-Student-t conditional-variance series. 5. Run python code/data_prep/reference_points_v3.py for the regime decomposition and reference-point construction. 6. Run python code/data_prep/unit_root_tests.py for the ADF, Phillips–Perron, KPSS, and Zivot–Andrews diagnostics. 7. Run python code/model/bayesian_main_spec.py for the main Bayesian estimation (4 chains × 4000 iterations). 8. Run python code/model/bayesian_dol_interaction.py for the RB6 dollarization-interaction specification. 9. Run python code/model/h4_loo_comparison.py for the LOO-CV nested-model comparison test.