Paper summary: Gaussian GenAI — Synthetic Market Data Generation

A GMM with K components over an n-dimensional observation vector x (e.g. x might be the 15 points on a yield curve) is:

p(x) = sum_{k=1}^{K}  w_k * N(x ; mu_k, Sigma_k)

w_k >= 0,  sum(w_k) = 1
mu_k    : mean vector   (dimension n)
Sigma_k : covariance    (n × n, symmetric positive-definite)

Think of each component as capturing one "market regime": a steepening environment, a parallel shift, an inversion. The mixture weights w_k say how often each regime occurs. K is a small integer — 3 to 7 in the paper's experiments — not a deep neural network with millions of parameters.

Analogy for QuantLib developers: it is similar to using a PiecewiseYieldCurve with K knot points but for probability distributions rather than discount factors. Each Gaussian component is a knot; EM is the bootstrap.

EM iterates two steps:

1. E-step (assign): given the current parameters, compute the posterior probability that each historical observation x_t belongs to each component k. This is a closed-form Bayes update.
2. M-step (refit): update mu_k, Sigma_k, and w_k to maximise the likelihood of the data given those assignments. Also closed form.

Fitting converges in seconds or minutes on daily financial time series of a few years. Compare to GAN training, which requires GPU hours and careful tuning of adversarial objectives.

Once the mixture is fitted, two operations are analytic:

• Marginal: drop dimensions from x to get the distribution of a subset (e.g. just the 5Y and 10Y swap rates). Each Sigma_k marginalises block-algebraically.
• Conditional: fix some dimensions (e.g. the overnight rate is 4.25%) and get the distribution over the remaining dimensions. This follows from the standard Gaussian conditional formula applied component-by-component.

This is the key advantage over GANs and autoencoders: those architectures have no closed-form conditional. They handle conditioning by appending the conditioning variable as an extra input feature, which does not produce a true conditional distribution. GMMs give you a genuine conditional from Bayes' theorem.

In ORE terms: you can pin the OIS discount curve and draw consistent swaption-vol realisations from the conditional, just as a QuantLib CalibratedModel fits to a given term structure.

Draw a synthetic observation in two steps:

1. Sample a component index k ~ Categorical(w_1, …, w_K) using a uniform variate.
2. Sample x ~ N(mu_k, Sigma_k) using standard Gaussian variates.

No neural network forward pass, no rejection sampling, no SDE simulation. Generation is as fast as drawing from a multivariate normal.

Lower K = simpler regimes. Vol surfaces are "smoother" in distribution space than overnight rates, hence fewer components.

Treat the missing dates as unobserved dimensions. Condition on the observed dates to get the posterior distribution of the missing values and sample from it. Because the conditional is analytic, there is no need for iterative MCMC or data-augmentation tricks.

• Real-world measure only. The GMM models P-measure dynamics. It does not enforce risk-neutral no-arbitrage constraints (e.g. calendar-spread monotonicity of option prices, butterfly positivity, put-call parity). Use it upstream of an arbitrage-free interpolation layer; do not feed raw GMM samples directly into QuantLib pricers without checking arbitrage.
• Stationarity (implicit). EM fits to the empirical distribution of the input time series without regard to time order. If the distribution has shifted over the sample period (e.g. a post-2022 rate regime versus pre-2022 near-zero rates), the fitted mixture averages across regimes rather than adapting. The practitioner must choose training windows carefully or apply differencing / log-returns before fitting.
• Multivariate normality of components. Each mixture component is Gaussian: heavy-tailed events (e.g. March 2020 vol spike) are modelled only by pulling the Gaussian means and covariances, not by fat-tailed component distributions. K may need to be large to capture extremes well.
• Daily granularity. The method is designed for end-of-day snapshots. Intra-day tick data would require an impractically large K to capture the microstructure, and tractability degrades.
• Fixed dimension n. All observations must lie in the same space. If the yield curve changes tenor points (e.g. a new 20Y benchmark is added) the model must be re-fitted from scratch.

• High-frequency / tick data is infeasible. Explicitly stated in the paper. Too many Gaussians would be needed; the model loses tractability.
• Small K overfits smooth data; large K overfits noisy data. Model selection (choosing K) requires cross-validation or information criteria (BIC/AIC). The paper reports empirically chosen K values but does not provide an automated selection recipe.
• No temporal structure. The GMM models the marginal distribution of one daily snapshot. It does not capture autocorrelation or mean-reversion across dates. Generating a time series path requires an additional temporal model (e.g. fit a first-order autoregression on the latent component assignments, or use GMM on log-returns and integrate).
• No arbitrage enforcement. Samples from the marginal or conditional may violate calendar-spread or butterfly constraints on vol surfaces. A post-processing arbitrage-removal step (see the companion paper on mixture-preserving, arbitrage-free vol-surface interpolation) is needed before using generated surfaces in QuantLib pricers.
• Covariance matrix size grows as n². A yield curve with 20 tenor points has a 20×20 covariance per component. With K = 7 components that is 2,800 free parameters. For a full swaption cube (say 10 expiries × 10 tenors × 5 strikes = 500 dimensions) this becomes intractable without dimensionality reduction (e.g. PCA preprocessing) first.
• Outperforms GANs/autoencoders only on short datasets (1–4 years daily). Deeper architectures may catch up with more data.