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

Takeaway. VaR is a quantile of losses. This paper does not fit a quantile net on returns; it trains generative, regime-switching networks, Monte-Carlos the paths, then reads a VaR threshold. Three design knobs matter: how you initialise, what the loss rewards, and how much history you feed.

Neural Networks and Value at Risk (Alexander Arimond, Damian Borth, Andreas Hoepner, Michael Klawunn, Stefan Weisheit, 2020; arXiv:2005.01686). Equities and long bonds, global / US / euro area / UK, up to about 1,250 weeks ending August 2018.

Fit for learners: Regime-switching nets for VaR: HMM initialisation, a two-objective loss (accuracy + realistic bull/bear mix), and a warning that cutting the sample hurts.

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

The problem

A Hidden Markov Model (HMM) is the textbook two-regime story (bull vs bear). Neural nets can generate more flexible paths, but random seeds, a single “be accurate” loss, and a thin sample can make VaR breaches worse than a boring benchmark.

The idea / core intuition

• Treat the net as a generator of returns inside a regime-switching wrapper; simulate; take the VaR quantile of simulated losses.
• Init: start weights from an estimated HMM instead of pure random seeds — fewer breaches in their tests.
• Loss: add a second objective so the implied bull/bear frequencies stay realistic, not just point-forecast accuracy. The balanced-incentive RNN is their winner.
• Data: cut training from 2,000 days to 1,000 and performance drops — they flag data hunger as a weakness.

A risk-metric bake-off—not a trading book.

Why after-work Python learners care

You already meet VaR as “the 5% worst return.” This paper is a checklist of training choices that change breach counts. Read it before stacking LSTMs on a 3-year ETF series and calling it risk.

Tiny evening exercise: on toy returns, compute historical 5% VaR with np.quantile(r, 0.05) and count how often later returns go below it. Learning only.

One-evening reading checklist

1. Generator + Monte Carlo vs direct quantile regression.
2. Why HMM initialisation beat random seeding on breaches.
3. What the second loss term (regime frequencies) is doing.
4. The 1,000-day vs 2,000-day warning.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. Academic VaR experiments are not a recommendation to size positions or satisfy a regulator.

What this is not

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

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—quantiles and loops are enough to read a VaR table; then open the PDF.