Home / Research
Reading Room
The argument, shown — with its own counter-arguments.
Editorial instrument: cream, 720px measure, every figure in mono, every claim shadowed by its limitation. Written for the reader who distrusts anything that cannot be broken.
Papers
Current writing.
8 min · methodology
The Practical Objection
"Tail hedging is a tax on returns." We answer it with the carry math.
Read12 min · technique
Quantitative Technics
Why diffusion beats Gaussian VaR in the only regime that matters.
Read10 min · perspective
The LLM-Model Perspective
Where language models help reason about tail risk — and where they must not.
ReadYou build systems. You assume every AI product is a wrapper around a public API until the source proves otherwise. Fair. Here is the source.
The source
Why should you trust a model you can’t see inside?
You shouldn’t — not on assertion. So we publish the math, the data provenance, and the rate at which the engine is wrong. The four papers below are the load-bearing structure: peer-reviewed, citable, and chosen because each maps to a specific failure mode in a concentrated book.
Read them as a spec sheet, not a sales deck. If a number looks soft, the citation tells you where to break it. The full estimator stack — the diffusion sampler, the GPD signal, the conditional-fidelity checks — is laid out in the quantitative techniques behind the engine.
Papers
The four papers behind the engine
Do diffusion models actually capture fat tails better than Monte Carlo?
Yes — and the gap is measurable. Fukunishi, Qiu & Takahashi (2026), Generating Synthetic Stock Return Distributions with Diffusion Models report that their tail-aware diffusion model recovers an S&P 500 return kurtosis of 13.38 against the real 16.70, while a Gaussian/Geometric-Brownian baseline collapses to 2.90 — the exact regime where standard VaR pretends crashes can’t happen. This is the engine’s core read on how tail risk is measured and offloaded for a concentrated holder.
What does an AI-routed hedge actually return for a concentrated holder?
Chou, Lin & Pham (2025), AI-fused construction portfolio investment system with risk hedging (Applied Soft Computing) built an adaptive risk-hedging long-short system that posted a 23.11% return at a Sharpe ratio of ≈ 3.0, holding stable returns and low volatility through market turbulence — the property a single-ticker position lacks by construction.
Is “tail risk” a priced factor or a story?
Priced, and persistently. Almeida, Ardison, Garcia & Vicente (2017), Nonparametric Tail Risk, Stock Returns, and the Macroeconomy (Journal of Financial Econometrics) sort stocks by tail-risk sensitivity and find the long-short spread earns 3.06% per month, surviving a 2.40% alpha after the Fama-French and momentum factors. Tail exposure is compensated — which means yours can be measured and offloaded.
Does dynamic tail protection beat just holding through the drawdown?
It does, net of cost. Packham, Papenbrock, Schwendner & Woebbeking (2017), Tail-risk protection trading strategies (Quantitative Finance) show a GPD-signal protection strategy on 11 years of DAX futures lifting the Sharpe ratio to 0.4587 versus 0.3022 for static buy-and-hold, with a better worst drawdown — and risk-averse investors paying a positive premium to switch.
Where it fails
Where the models fail
The engine carries a published 4.2% false-comfort rate — the cases where it reports a tail smaller than the realized one.
Stated limits, up front:
- Diffusion is univariate first. Cross-asset joint tail dependence is the active boundary (Fukunishi et al. name it as their primary open problem). Multi-leg correlation breakdown is modeled, not solved.
- Bear regimes have fewer samples. Conditional fidelity degrades exactly where you care most; the engine flags low-sample regimes rather than smoothing over them.
- Historical support is finite. A regime with no precedent in ~20 years of data is extrapolation, and labeled as such in the output.
If your tail is smaller than you feared, the document says so. That is the point of stating the rate. We field the standard pushback — sample bias, overfitting, regime change — in the practical objections to the methodology.
The architecture
“Couldn’t an LLM just do this?”
No — and the reason is architectural, not branding. A generic LLM predicts the next plausible token; it has no native model of a return distribution’s tail mass, so it regresses to the consensus middle exactly where you need the extreme. Gaussian VaR fails for the opposite reason: it assumes thin tails, underestimating crash density beyond |r| > 5% (Fukunishi et al.).
A diffusion model learns the full distribution by reversing a noise process, so it allocates probability mass to the heavy tails directly — the difference between compiling 50,000 stress paths that include the crash and averaging your way into denying it. We unpack exactly where an LLM breaks on tail mass, and the live dashboard runs the diffusion engine against your own holding rather than a generic index.
3 fields · 48-hour document · no call, no sequence. If your tail is smaller than you feared, the document will say so.
Frequently asked questions
Is the tail-risk methodology academically validated?
Yes. It rests on four peer-reviewed papers — Fukunishi, Qiu & Takahashi (2026) on diffusion fat tails, Chou, Lin & Pham (2025) on AI-fused hedging, Almeida et al. (2017) on tail risk as a priced factor, and Packham et al. (2017) on protection trading. Each citation is named above and maps to a specific failure mode, not a marketing claim.
Why use a diffusion model instead of Monte Carlo?
Because Monte Carlo built on Gaussian assumptions denies crash density beyond |r| > 5%, collapsing S&P 500 kurtosis to 2.90 against the real 16.70. A diffusion model learns the full distribution and recovers a kurtosis of 13.38 — close to reality — so it places probability mass on the heavy tails Monte Carlo prices out.
What is the engine’s failure rate?
The engine carries a published 4.2% false-comfort rate: cases where it reports a tail smaller than the realized one. The rate is stated in the output, alongside its known boundaries — univariate-first diffusion, sparse bear-regime samples, and finite historical support that is flagged as extrapolation.
Can I see the underlying math?
Yes. The four citations are listed in full above, and the estimator stack — diffusion sampler, GPD signal, and conditional-fidelity checks — is set out in the quantitative techniques. Nothing in the engine is asserted without a breakable number behind it.
Is this just an LLM wrapper?
No. A generic LLM predicts the next plausible token and regresses to the consensus middle exactly where the tail matters; it has no native model of return-distribution tail mass. The engine is a diffusion model that reverses a noise process to allocate mass to the extremes — the architectural difference is detailed in the LLM model perspective.