When a stock-prediction model trained on one market fails on another, the usual explanation is that the inputs changed: volatility, liquidity or volume look different. There is a second explanation that finance rarely tests. The same input pattern may be followed by different outcomes in different markets, because local rules and participants "label" that pattern differently. A volume spike followed by a gap might mean continuation in a retail-heavy market with daily price limits and reversal in an institution-dominated one. Medical imaging has a well-developed toolkit for this situation. There, several experts trace the same scan and draw different boundaries, a problem called inter-observer variability. Models such as Pionono learn each expert's systematic bias and personal inconsistency as a small probability distribution, and they predict what a given expert would say.
We propose to treat each market as a rater. A shared network reads the market state. Each market gets a learned distribution over a low-dimensional "style" vector. The mean of that distribution captures how the market systematically maps patterns to outcomes. Its spread captures how inconsistently it does so. A new market is handled the way the Latent Doctor Model handles a new pathologist: its style is inferred from a short support set of its own recent labeled history, with no retraining. The output is a forecaster for unseen markets and, more importantly, a measurement. For any pair of markets, the method estimates how much of the transfer gap comes from input shift and how much from disagreement about outcomes. The style spread also gives a per-market confidence signal that can size positions or trigger abstention.
Core Claim
Cross-market failure contains a substantial component of conditional disagreement: the same features, different outcomes. That component can be represented as a per-market latent distribution and inferred from a few months of target data. Doing so should close more of the transfer gap than input-side fixes such as per-market normalization. It should also do so with less target data than fine-tuning. The variance of the latent should track where the model will be wrong in a new market. If per-market normalization alone closes the gap, the claim is false and finance's input-drift framing is right. That is a useful negative result, and the experiments below are designed to detect it.
Where This Sits
The finance side supplies the motivation and the benchmarks. Implicit-Causality-Exploration-Enabled Graph Neural Network for Stock Prediction builds a Granger-causal stock graph for 600 Chinese equities. Its ablation found that media-sentiment relations hurt accuracy, which the authors read as a retail-market idiosyncrasy. That is exactly the kind of market-specific mapping a rater latent should absorb. The paper never tests other markets, and neither does A Learnable Wavelet Transformer for Long-Short Equity Trading, which learns frequency filters and trains directly on return and Sharpe objectives over U.S. hourly data. We reuse its lesson that the training objective should reflect trading outcomes. We do not reuse its single-market evaluation. HiveMind treats disagreement among trading agents as a defect to be prompt-tuned away. We take the opposite stance, as the medical literature does: persistent disagreement is structure worth modeling.
The medical side supplies the method. Probabilistic Modeling of Inter- and Intra-observer Variability gives each rater a Gaussian in an 8-dimensional latent space, whose mean is bias and whose covariance is inconsistency, trained with a variational objective. It also validates on an external institution (Arvaniti TMA), which is the protocol we copy across markets. The Latent Doctor Model replaces fixed rater identities with an encoder over a support set of the rater's other labeled cases. It shows that predictive entropy tracks real panel disagreement, with Spearman correlation up to 0.728 on 20-pathologist ISUP grading. We need that support-set encoder because new markets have no identity embedding.
Both medical models work on images with many raters labeling the same item. Financial markets never label identical items. We handle this by defining "same input" in a shared, market-normalized feature space and testing the assumption explicitly. The Diagnosis-First work, Calibrate the inter-observer segmentation uncertainty via diagnosis-first principle, weights raters by downstream diagnostic utility. Its known weakness is that the learned weights depend on the downstream classifier. That weakness argues against making utility-weighted consensus our core idea, because it would overfit on short financial windows.
Model
Let $x$ be a market-normalized feature window for one stock-day: OHLCV over the past $L$ days, z-scored within market. Let $y$ be the next-period direction or return, $m$ the market, and $S_m$ a support set of $(x, y)$ pairs from market $m$'s recent history. The predictor is
$$p(y \mid x, S_m) = \int p_\theta\big(y \mid f_\phi(x), z\big)\, q_\psi(z \mid S_m)\, dz, \qquad q_\psi(z \mid S_m) = \mathcal{N}\big(\mu_\psi(S_m), \Sigma_\psi(S_m)\big).$$
Here $f_\phi$ is a shared encoder. We use a temporal backbone, with a DSF-GNN-style relational variant as a second backbone. $p_\theta$ is a head conditioned on the style vector $z$, and $q_\psi$ is a permutation-invariant set encoder. Training is episodic over source markets and source years. Each episode samples a market-period, a support set and disjoint query days. The loss is the variational lower bound plus a differentiable Sharpe term on the query days, following the wavelet-transformer paper's finding that pure point-wise losses misalign with trading.
To make the gap decomposition precise, define the target-market loss of three models. $\mathcal{L}_{\text{src}}$ is the source model applied directly. $\mathcal{L}_{\text{in}}$ is the same model after input-side adaptation, meaning target-specific normalization plus importance reweighting. $\mathcal{L}_{z}$ is the source model given only the inferred target latent. The share of the transfer gap attributable to conditional disagreement is
$$\rho = \frac{\mathcal{L}_{\text{src}} - \mathcal{L}_{z}}{\mathcal{L}_{\text{src}} - \mathcal{L}_{\text{oracle}}},$$
where $\mathcal{L}_{\text{oracle}}$ is a model trained on the target market's full history. The analogous ratio with $\mathcal{L}_{\text{in}}$ gives the input-shift share. Reporting both, with the remaining share left unexplained, is the central empirical product.
Research Program
We first assemble a multi-market daily panel using the exact feature set and filtering style of the DSF-GNN study. The anchor is the Tushare China 600-stock universe (January 2019 to January 2022). We add U.S. equities filtered by the same rules over the same window, and further held-out exchanges that differ in daily price limits and short-selling rules. Using the same window and features keeps calendar effects from masquerading as market effects. Each market is additionally split by year, so that "market as rater" can be compared with "year as rater." That separates cross-market from cross-regime disagreement.
Before touching real markets, we build a controlled testbed. We simulate several synthetic markets sharing one feature process but with known differences in outcome mapping. One truncates returns at a price limit. One adds a momentum or reversal tilt as a function of a single feature. One only rescales input volatility. The decomposition ratio must recover the known answer: near zero conditional share for the pure input-rescaled market and high for the outcome-tilted ones. If it cannot, it is not trustworthy on real data.
On real data we then answer three questions. First, does support-set style inference transfer to a held-out market better than input-side adaptation? Second, how does performance scale with support-set length, from two weeks to a year? Third, does the inferred style spread predict where the model will be wrong? For the last question we follow the Latent Doctor Model's evaluation and correlate predicted spread with realized error dispersion across stocks and months. Finally, we probe interpretability, a weakness both Pionono and DiFF leave open. Do the latent means of price-limited markets separate from unlimited ones along a consistent direction, and does moving a U.S. latent along that direction reproduce limit-like behavior in the head's predictions?
Baselines
The naive reference point is the source-trained model applied unchanged to the target market. Beside it sit an always-majority-direction predictor and a buy-and-hold portfolio. All are run on identical target test windows. Any proposal that does not beat the unchanged source model by a margin exceeding seed-to-seed variation has no claim.
The strongest in-market method in the file is DSF-GNN, which reports accuracy of 54.97% and Matthews correlation of 0.0821 on the Tushare China set (Implicit-Causality-Exploration-Enabled Graph Neural Network for Stock Prediction). We reproduce it on our China split and train it from scratch on each target market's full history. It then serves as the oracle in the decomposition ratio. A proposal using only a short support set will not be expected to match it. Getting within its seed variance using a few months of target data would count as a strong result. For cross-market transfer of any of these models, no reported result was found on Lacuna, so all transfer numbers will be produced here. For the trading metric, the wavelet transformer reports a Sharpe ratio of 2.157 and cumulative ROI of 0.607 on frictionless U.S. hourly data (A Learnable Wavelet Transformer). Because that number omits transaction costs and is single-market, we use its architecture as an alternative backbone rather than as a target number. We report all Sharpe ratios both with and without costs.
The strongest generic competitor is a pooled model with a learned point embedding per market, adapted to a new market by fine-tuning on the same support set the proposal sees. It gets the same backbone, the same optimizer budget, and a hyperparameter search of equal size on source-market validation folds. Beating it requires either higher Matthews correlation and cost-adjusted Sharpe at matched support length, or equal performance with a usable uncertainty signal it lacks.
A per-market ensemble, with one model per source market averaged at test time, is the analogue of the deep-ensemble baseline the Latent Doctor Model compares against. Its member disagreement is the natural competing uncertainty signal.
Two ablations isolate the key ingredients. The first replaces the Gaussian latent with a deterministic point vector, which removes the inconsistency term and tests whether the spread earns its place. The second replaces the support-set encoder with fixed market identities plus fine-tuning, which tests whether amortized inference matters.
Why not just…?
Why not just fine-tune a pooled model on the target market? This is the real threat. With enough target data, fine-tuning should match or beat any amortized inference. The proposal therefore stakes its case on the low-data regime and on the uncertainty signal, not on asymptotic accuracy. The support-length sweep settles it. If fine-tuning matches the latent model even at two to eight weeks of support and its predictive entropy correlates with errors as well as the latent spread does, the probabilistic machinery is unnecessary and we will say so. The decomposition ratio remains useful in that case, because it can be computed with fine-tuned models too.
Why not just normalize inputs per market? If cross-market failure were pure input drift, as the finance literature implicitly assumes, z-scoring and importance reweighting would close most of the gap. This is precisely the input-shift arm of the decomposition, so it is a baseline and a hypothesis at once. The synthetic testbed shows whether our measurement can tell the two cases apart. A large input-shift share on real markets would refute the core claim.
Why not just train on the target market alone? Short target histories are noisy. The confidence interval spanning zero over 36 months in Leakage-Aware Benchmarking of LLM Forecasting shows how quickly financial evaluation becomes underpowered. Target-only training at each support length is run as a baseline. We expect it to lose at short lengths and converge to the oracle at long ones. The crossover point tells practitioners how much history makes transfer pointless.
Why not just use ensemble disagreement for uncertainty? Ensemble variance captures uncertainty about model parameters. The style spread is meant to capture how inconsistently a market maps patterns to outcomes, which is the intra-observer term in Pionono. These can coincide. We compare both signals on the same error-dispersion correlation and abstention curves. If the ensemble is as good, the case for the inconsistency term rests only on its compute advantage, which the Latent Doctor Model already documents in medicine.
Proposed Experiments
The decisive check comes first: leave-one-market-out evaluation on the multi-market panel. We report accuracy, Matthews correlation, and cost-adjusted long-short Sharpe at support lengths of two weeks, two months, six months and one year, averaged over at least five seeds, with block-bootstrap intervals over months.
The proposal succeeds if three things hold. The latent model beats both the unchanged source model and the fine-tuned pooled model at the two shortest support lengths. The conditional share $\rho$ is substantially above the input-shift share for structurally different market pairs, such as price-limited versus unlimited, and near zero for similar pairs. The Spearman correlation between latent spread and realized error dispersion exceeds that of ensemble disagreement.
A second check repeats the analysis with years as raters within one market. If cross-year conditional shares rival cross-market ones, the "market" framing is really a "regime" framing, and the paper will be written that way.
A third check, run first in time, is the synthetic testbed recovering known decomposition shares. Real-market numbers are reported only if it passes.
All experiments use walk-forward splits with an embargo between support and query periods, following the leakage discipline of the macro-factor benchmark. This keeps support-set inference from seeing the future it is scored on.
References
- Implicit-Causality-Exploration-Enabled Graph Neural Network for Stock Prediction
- A Learnable Wavelet Transformer for Long-Short Equity Trading and Risk-Adjusted Return Optimization
- HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Leakage-Aware Benchmarking of LLM Forecasting: Real-Time Nowcasts as the Decision-Time Input for Macro Factor Ranking