bayesianOU

Bayesian Nonlinear Ornstein-Uhlenbeck Models with Stochastic Volatility

Overview

bayesianOU fits a Bayesian nonlinear Ornstein-Uhlenbeck (OU) / nonlinear error-correction model with cubic drift, stochastic volatility (SV), and Student-t innovations. Sector-specific parameters use a non-centered hierarchical (partial-pooling) parameterization, and sampling is done with Stan (reduce_sum for within-chain parallelism). Model comparison uses PSIS-LOO (Vehtari, Gelman, Gabry 2017).

Installation

# install.packages("remotes")
remotes::install_github("IsadoreNabi/bayesianOU")

# Stan backend (cmdstanr recommended):
install.packages("cmdstanr", repos = c("https://mc-stan.org/r-packages/", getOption("repos")))
cmdstanr::install_cmdstan()
# Or rstan:
install.packages("rstan")

Quick Start

library(bayesianOU)

Y   <- as.matrix(your_prices_data)             # market prices by sector
X   <- as.matrix(your_production_prices_data)  # prices of production
TMG <- your_tmg_series                         # aggregate profit rate
COM <- as.matrix(your_com_data)                # organic composition of capital
K   <- as.matrix(your_capital_data)            # total capital by sector

results <- fit_ou_nonlinear_tmg(
  results_robust = list(),
  Y = Y, X = X, TMG = TMG, COM = COM, CAPITAL_TOTAL = K,
  fit_window = "train",   # honest out-of-sample evaluation
  chains = 4, iter = 4000, warmup = 2000,
  verbose = TRUE
)

validate_ou_fit(results)            # MCMC + LOO (incl. Pareto-k) + OOS
kappa_stability_evidence(results)   # dynamic mean-reversion evidence
plot_beta_tmg(results)
plot_drift_curves(results)

Model Specification

For each sector s, on training-standardized series, the one-step (Euler, dt = 1) increment is

dY[t,s] = kappa_s * (theta_s - Y[t-1,s] + a3_s * (Y[t-1,s] - theta_s)^3)   # cubic OU drift
        + (beta0_s + beta1 * zTMG[t]) * X[t-1,s]                           # TMG-modulated pass-through
        + gamma * COM_std[t-1,s]                                            # organic composition in mean
        + eps[t,s],     eps ~ Student-t(nu, 0, sigma[t,s])
log sigma^2[t,s] = h[t,s],   h ~ stationary AR(1)  (alpha_s, rho_s, sigma_eta_s)

• kappa_s mean-reversion speed, theta_s equilibrium level, a3_s < 0 cubic term. The implied long-run equilibrium is Y* = theta_s + (beta/kappa_s) X, i.e. a nonlinear error-correction toward a linear function of the prices of production.
• theta_s, kappa_s, a3_s, beta0_s are built hierarchically as intercept + slope * COM_s + sd * z, z ~ N(0,1) (non-centered).
• nu > 2 (finite variance). sigma[t,s] is the Student-t scale; the conditional standard deviation is sigma * sqrt(nu/(nu-2)).

Methodology and validity (read before reporting results)

Training/test and “out-of-sample”. A Bayesian fit on all data is perfectly valid for inference and for PSIS-LOO (which approximates leave-one-out predictive performance from the full-data posterior — this is exactly what rstanarm does, and it is not leakage). Leakage only arises if one labels a block “out-of-sample” while it still entered the likelihood. This package keeps the two designs coherent via fit_window:

• fit_window = "train" (default): the likelihood and log_lik are summed only over the training window 2:T_train; the test block is genuinely held out, so evaluate_oos is a real out-of-sample evaluation.
• fit_window = "full": fit on all observations (full-information). PSIS-LOO is then valid over all points, but evaluate_oos becomes in-sample.

Priors are configurable and neutral by default. The key hypothesis prior is neutral: beta1 ~ Normal(0, 0.5) (no sign baked in). The Student-t df uses a weakly-informative nu_tilde ~ Gamma(2, 0.1) (prior mean nu ~ 22), and SV persistence rho_s ~ Normal(0.7, 0.2). Override any of these through the priors argument and run a sensitivity analysis.

Leakage in preprocessing is avoided by default. use_train_loadings = TRUE computes the common-factor loadings from the training window only and then projects the full series, so the orthogonalized TMG regressor does not see the future.

Stability assumption. a3_s < 0 is enforced, so mean reversion strengthens with the deviation. This is an assumption (global stabilization), not an estimated result; it precludes detecting locally expansive / self-amplifying regimes.

Half-life. With the Euler dt = 1 discretization, the discrete persistence of the linear part is 1 - kappa; interpret half-lives as log(0.5)/log(1 - kappa), not log(2)/kappa (these coincide only as kappa -> 0).

PSIS-LOO caveat. The model has one latent SV state per observation, so plain PSIS-LOO tends to be optimistic and to produce high Pareto-k. validate_ou_fit surfaces a Pareto-k summary and warns; for genuine forecasting assessment prefer leave-future-out (LFO-CV).

Validation layers

• Internal: R-hat, ESS, divergences, Pareto-k (validate_ou_fit).
• Synthetic recovery: simulate from the generative process with known parameters, fit, and check coverage (see tests/testthat/test-recovery.R).
• Organic / external: held-out metrics (fit_window = "train") and PSIS-LOO comparison against baselines.

References

• Vehtari, Gelman, Gabry (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing.
• Bürkner, Gabry, Vehtari (2020). Approximate leave-future-out cross-validation for Bayesian time series models. J. Stat. Comput. Simul.
• Gelman (2006). Prior distributions for variance parameters in hierarchical models. Bayesian Analysis.

Citation

@software{bayesianOU,
  author = {Gómez Julián, José Mauricio},
  title  = {bayesianOU: Bayesian Nonlinear Ornstein-Uhlenbeck Models},
  year   = {2024},
  note   = {ORCID: 0009-0000-2412-3150},
  url    = {https://github.com/IsadoreNabi/bayesianOU}
}

License

MIT License