The Paranormal Distribution: When Markets Live Outside the Bell Curve
By Dimitrios Thomakos
Update (18/3/2026): Kostas Nikolopoulos has alerted me that the
paranormal distribution has been already there in this
post: https://mathsrant.com/2025/10/31/on-the-statistical-properties-of-the-paranormal-distribution/.
I happily attribute the name "Paranormal Distribution" and original idea
to Shub Das but will keep the novel mathematical content of the post. If
the working paper is ever send for publication a formal attribution will
be made.
Real-world data — inflation rates, GDP growth, stock returns, Bitcoin prices — rarely behave the way textbooks assume. The familiar bell curve assigns vanishingly small probabilities to the events that matter most: market crashes, inflation surges, post-recession bouncebacks. Our new paper introduces (in a funny but constructive way) the Paranormal distribution, a two-component mixture of skew-t distributions that captures what standard models will miss: the coexistence of distinct economic regimes, each with its own direction, volatility, and tail thickness. Think of it as a principled way of saying that markets do not live in one world — they switch between states, and the shape of risk is fundamentally different in each. We derive the distribution's theoretical properties, identifiability conditions, and develop a differential evolution estimator with k-means initialisation that reliably navigates the rugged likelihood landscape these mixtures produce.
Fitting the Paranormal to five empirical series — US CPI inflation, real GDP growth, SPY, the 3× leveraged TNA ETF, and Bitcoin — the results tell a coherent story. Bitcoin is where the Paranormal wins most decisively: it outperforms the next-best model (the Normal-Inverse Gaussian) by ΔAIC = 31 units, uncovering a bull-accumulation regime with high volatility and a bear-consolidation regime that is paradoxically calmer — the inverse of what equity markets exhibit. For SPY and CPI the Paranormal finishes a statistically negligible second (ΔAIC < 0.5), confirming its competitiveness without overstating the case. The one honest caveat is small samples: with only 103 quarterly GDP observations the nine-parameter model is penalised by AIC despite achieving the highest log-likelihood of all competitors — a reminder that richer descriptions require richer data. Across all five series the Paranormal never finishes lower than second, and the Q-Q plots and tail-exceedance charts confirm that its mixture structure captures the asymmetric, heavy-tailed regimes that define modern financial and macroeconomic risk.
