Real or generated? S&P 500 returns
Each card shows two series of 250 daily returns of the S&P 500 stock index, about one trading year. One is real, from the years 2005–2019; the other was produced by the generative adversarial network (GAN) of Lecture 8, trained on the same years. Pick the real one.
The model you play against is not a network but a statistical check from the lecture: volatility clustering. For each series it computes the average autocorrelation of the absolute returns over lags 1 to 5 (one trading week) and calls the series with the larger value real.
What the game shows
- The GAN learned the distribution of a single day well: its 1% Value-at-Risk (VaR) is −3.61% against −3.48% for the real returns, and its tails are fat (excess kurtosis 6.65 against 11.86).
- It did not learn the order of the days: the autocorrelation of the absolute returns at lag 1 is 0.28 for the real returns and −0.01 for the GAN. Real markets have calm and turbulent periods; the GAN scatters its large moves at random.
- This one statistic picks the real series in 98.2% of 1,000 random pairs. A generator for stress tests must produce whole paths, not single days: this is why sequence models and diffusion models follow in the lecture.
Data: S&P 500, daily adjusted closing prices (Yahoo Finance), log returns 4 January 2005 – 31 December 2019; GAN samples: lectures/L8/cache/gan.npz (generator of seed 1). The 24 pairs of the game are drawn at random (seed 20261108); every game shows 10 of them, in random order and position. Code: lectures/L8/make_game_L8.py.

