Mode Collapse in GANs

A generative adversarial network trains a generator to produce fake samples and a discriminator to tell fake from real, with the generator improving by learning to fool the discriminator. In theory the generator should learn to reproduce the full diversity of the real data — every "mode," or distinct cluster of outcomes, the real distribution contains. In practice, generators frequently find a shortcut: produce a small handful of outputs that reliably fool the current discriminator, and keep producing essentially the same thing over and over, ignoring the rest of the real distribution's diversity entirely. That collapse onto a narrow subset of outputs is mode collapse.

It happens because the generator only ever gets a training signal from whatever currently fools the discriminator — if one particular kind of output is working, the generator has every incentive to keep producing it rather than risk diversifying into outputs the discriminator might catch. The discriminator eventually catches on and starts rejecting that narrow output too, at which point the generator can lurch to a different narrow mode rather than actually learning to cover the full distribution — a cycle that can repeat without ever converging to real diversity.

For a GAN trained to generate synthetic financial return paths, mode collapse would show up as the model generating a suspiciously narrow range of scenarios — say, always mildly upward-trending paths with a similar volatility profile — even though the real historical data contains sharp crashes, calm range-bound periods, and momentum rallies. A synthetic-data set with collapsed modes looks fine on summary statistics like the mean, but understates tail risk and scenario diversity, which is exactly the property synthetic data is often generated to preserve.

What this means in practice

Mode collapse is why GAN-generated synthetic market data has to be checked for diversity specifically — comparing the range and clustering of generated scenarios against the real data, not just matching moments like mean and variance — before it's trusted for stress testing or scenario generation. It's also the main reason alternatives like Wasserstein GANs, which use a different training objective specifically designed to resist this collapse, are often preferred over a vanilla GAN for financial applications where tail scenarios matter.