Introducing Neural Networks and Value at Risk for after-work quants

Takeaway. VaR is a quantile of simulated paths, not a Keras “predict tomorrow’s return.” This paper generates returns from a regime-switching net (bull vs bear), then reads the tail. Three design knobs: how you initialise, what the loss rewards, and how much history you feed.

Neural Networks and Value at Risk (Arimond et al., 2020; arXiv:2005.01686). Equity indices and long bonds, global / US / euro area / UK, up to ~1,250 weeks ending August 2018.

Fit for learners: Regime-switching generative nets for VaR: HMM init, balanced bull/bear loss, and why 1,000 vs 2,000 days of data matter.

This Hub note is a plain-language intro. Not investment advice.

The problem

A 99% weekly VaR is a rare-event number. Parametric GARCH-plus-normal is simple but misspecified. Hidden Markov models (HMM) already think in regimes. Neural nets can generate richer paths—but they overfit, ignore realistic bull/bear frequencies, and collapse when you halve the sample.

The idea / core intuition

• Simulate asset returns in a generative regime-switching setup; VaR is the empirical threshold of those paths.
• Init: random seeds vs starting from a fitted HMM—HMM-init cuts VaR breaches (realised return worse than the estimated threshold).
• Loss: add a second objective so the net stays in empirically realistic bull vs bear frequencies, not just point accuracy. A balanced-incentive RNN beats single-incentive RNNs and classic benchmarks.
• Data: cut training from 2,000 to 1,000 days—performance drops. Nets are data hungry.

A risk-lab comparison—not a desk VaR you should ship.

Why after-work Python learners care

You learn to separate generator, loss, and evaluation. “Did we breach?” is a count; “was the regime mix realistic?” is a second score. That is closer to how risk teams argue than a single MSE.

Tiny evening exercise: from a list of weekly returns, print the 5% quantile with sorted(r)[int(0.05*len(r))]. Then count how often a new week is worse. Learning only.

One-evening reading checklist

1. VaR here is a Monte-Carlo threshold, not a closed-form σ × z.
2. Why HMM initialisation beats random seeds on breach frequency.
3. What the second loss term (regime frequencies) is for.
4. What happens when you train on half the days.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. Arimond et al. is academic risk research—not a recommendation to set bank or fund VaR limits from a neural net.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not a risk-product pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—sorting a list and picking an index is how a historical quantile starts.