Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes

arXiv:2607.27188 · cs.LG, q-fin.CP, q-fin.PR, q-fin.ST · Submitted 2026-07-29 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes".

Tom: The gist: Accurate option prices do not imply accurate recovery of the latent risk-neutral density.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So to wrap up this discussion on "Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes", what's the simple meaning for us listening right now?

Jane: Basically, you can get a very good price for an option based on sparse quotes, but that price might be misleading if you try to reverse-engineer the true underlying risk-neutral density.

Lu: The title itself points to the core challenge: inverting a problem where your observations are irregular option quotes, and the output is a latent risk-neutral density.

Meng: It means we can't just rely on one method—whether it's a specific AI architecture or a classical math formula—to get this density estimate reliably across the board.

Lalam: The paper shows that what works depends on what you are trying to evaluate, whether you want aggregate metrics, or if you want better performance in specific tail functionals under structural misspecification.

Tom: And the authors conclude that we need to be very careful about how we describe these inverse problems when talking about risk-neutral densities in the future.

Jane: They’re emphasizing that validation concerns observable prices versus simulator-truth densities, and whether you are looking at aggregate synthetic metrics or specific functional errors.

Lu: The evidence supports neither a universal learned estimator nor a universal classical one; the appropriate prior depends on the target.

Meng: It’s about being honest with your setup: what you supervise, what structure you assume, and what functional you actually care about downstream.

Lalam: The matched representation experiments further show that density output alone is not sufficient; representation, supervision, architecture, and postprocessing all jointly determine recovery quality.

Tom: So the main point is that in this inverse problem research on irregular quotes, there isn't one single answer for the best way to learn a latent density from prices.

Conclusion: Tom: So we're wrapping up this look at "Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes." It’s about trying to figure out a secret risk density just by looking at messy option prices.

Jane: That’s right, and the authors are showing that getting an accurate price isn't the same as accurately recovering that hidden density structure.

Lu: They use two different sets of data to test this, one where they have perfect simulator truth densities and another using real market quotes from NIFTY. It shows how much the source of your information matters for the result.

Meng: From an engineering standpoint, it seems like the problem isn't just picking a better learning model; it’s about defining what kind of density you’re trying to recover in the first place.

Lalam: The main point is that there isn't one single winning architecture for this inverse problem. It really depends on your supervision target and how you describe the mathematical structure.

Tom: So, if I only listen to this, I should think that I can’t just plug in a fancy new AI model and expect it to magically give me the perfect risk density estimate every time.

Jane: Exactly. The paper suggests that we need a much clearer idea of what we’re looking for before we start training anything up.

Meng: It means the practical impact isn't just about getting a better price prediction; it's about building systems that understand which kind of errors—like tail risk versus average price error—we want to fix.

Lu: And if you look at the next part, they show that how you set up your supervision target and your structural assumptions really dictates whether a learned operator wins or not.

Tom: So the big picture is that this isn't just an incremental improvement; it’s about changing how we frame these kinds of inverse problems in finance and AI.

Jane: It suggests that the next step isn't just building bigger models, but defining better priors based on what you actually need to predict.

Meng: And that leads us right into the next piece of the puzzle—how do we actually define those targets when we’re dealing with real-world data?

Georgia Institute of Technology · Florida Institute of Technology

cs.LG, q-fin.CP, q-fin.PR, q-fin.ST

Submitted: 2026-07-29

Updated: 2026-10-12

Comments: 8 pages, 4 figures, 2 tables. Accepted to the 7th ACM International Conference on AI in Finance (ICAIF 2026)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: The gist: Accurate option prices do not imply accurate recovery of the latent risk-neutral density.

Key concepts

Latent Risk-Neutral Density: This is the true probability distribution of future asset prices under a risk-neutral measure. It is not directly observable in market quotes but is crucial for pricing derivatives accurately. The paper explores how different models attempt to recover this hidden distribution from imperfect market data.
Controlled Benchmark vs. Chronological Benchmark: The researchers used two types of tests: one where they had perfect simulator-truth densities for training, and another using actual historical NIFTY option prices that were held out. This separation helps determine if the models learn the true underlying density or just patterns in market prices.
Density MSE Supervision: This is a specific mathematical objective used to train the AI models. It measures how well the model's predicted probability distribution (density) matches a known, ground-truth distribution during training. The results show that simply minimizing this error isn't enough; other constraints matter.
Target-Dependent Inductive Bias: This means that the best learning approach (the 'bias' the model develops) is not universal. It changes based on what you are trying to achieve—whether it's matching aggregate prices, improving a specific functional like variance, or handling a particular market regime.

Terminology

Benchmark — 4× in this paper · explained in 2 episodes
DeepONet — 2× in this paper · explained in 2 episodes
Adaptation — 1× in this paper · explained in 1 episode
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist: Accurate option prices do not imply accurate recovery of the latent risk-neutral density.

Problem and Evaluation Principle

The study separates evidence by using two benchmarks: a controlled benchmark exposing simulator-truth densities for latent evaluation, and a chronological NIFTY benchmark testing only held-out market pricesThe simulator draws exact densities from martingale process families, prices continuous quotes analytically, and reveals the density only for controlled supervision and evaluationThe observed input is an unordered set Q = (ki, τi,Ce)ni=1, Ci = C(ki, τi) + εi, where both n and the locations vary by example

Methods and Experimental Design

The researchers compare four learned backbones with direct parametric and regularized density estimators, surface smoothers, and Breeden–Litzenberger (BL) extractionThe primary protocol uses only supervised density MSE, Lmain = Lpdf. (6)The objective ablation compares density MSE with combinations of observed-quote MSE, Eq. (5), a forward-moment penalty, and an H1-style grid roughness penaltyFor reporting, predicted prices are certified against bounds, strike monotonicity, vertical-spread slopes, physical-strike convexity, and calendar monotonicity

Learned Models and Benchmarks

Four learned backbones were compared with direct parametric and regularized density estimatorsDeepONet combines a mean-pooled quote branch with a coordinate trunkThe primary benchmark trains density outputs for all fourUnder density-MSE supervision, the quote transformer has the lowest stable learned L1 and W1, while DeepONet has the best learned fixed-tail, 1% quantile, variance, and price errorsThe direct mixture baseline is stronger on aggregate price, L1, W1, and fixed-tail error

Results and Conclusions

The conditioning experiment shows that after enforcing mass and forward constraints, 95 of 126 pricing directions are numerically null, and two densities separated by L1 = 0.061 produce identical prices on the covered strikesThe evidence supports target-dependent inductive bias, not a universal winnerThe relevant scientific result is not a single winning architecture: the inverse problem, supervision target, structural specification, and downstream functional jointly determine the rankingA well-specified mixture is strongest on aggregate synthetic metrics, while learned operators improve selected functionals and the misspecified Merton regimeThe chronological NIFTY study reinforces the separation between latent and observable evaluationTest-time adaptation substantially improves the amortized DeepONet, but per-expiry mixture and SVI fits remain more accurate on held-out market pricesThe relevant scientific result is not a single winning architecture: the inverse problem, supervision target, structural specification, and downstream functional jointly determine the rankingRND recovery claims should instead state the supervision available, the assumed distributional structure, the downstream functional of interest, and whether validation concerns observable prices or simulator-truth densitiesThe evidence supports neither a universal learned estimator nor a universal classical oneThe appropriate prior depends on the targetThe matched representation experiments further show that density output alone is not sufficient; representation, supervision, architecture, and postprocessing jointly determine recovery qualityThe chronological NIFTY study reinforces the separation between latent and observable evaluationThe evidence supports neither a universal learned estimator nor a universal classical one.

Improvements for AI systems

Bold header: Latent Density Recovery from Sparse Data

The improved system can perform Inverse Learning of Latent Risk-Neutral Densities from Irregular Option Quotes, allowing it to estimate an underlying risk-neutral density even when only sparse, irregular option quotes are available.

Bold header: Target-Dependent Inductive Bias Selection

The AI can select the most appropriate recovery strategy by leveraging the target-dependent inductive bias observed, meaning it will choose between a well-specified lognormal mixture for aggregate metrics or learned operators for specific functionals like the 1% quantile and variance error.

Bold header: Stress Testing and Misspecification Detection

The system can identify structural weaknesses by performing stress tests against different families, as the paper shows that learned operators improve selected tail functionals and recovery under structural misspecification, such as in the Merton family.

Bold header: Evaluation of Representation vs. Architecture

The AI can determine if a model's success stems from its internal structure by comparing Matched representation results, showing that density output alone is not sufficient; representation, supervision, architecture, and postprocessing jointly determine recovery quality.

Bold header: Market Data Adaptation Strategy

For real-world application on the NIFTY data benchmark, the system can employ test-time adaptation to reduce error by 28.3%, though the paper cautions that this does not close the gap to local classical refitting methods.