Towards Financial World Modeling

Abstract: Building a world model requires a state representation useful for planning and decision-making---potentially over tasks unknown at training time. In the context of financial markets, planning and decision-making may require a model to reason about market-wide conditions, asset-specific expected returns, liquidity, volatility, and cross-asset relationships. Yet financial representation learning has largely been evaluated on individual predictive tasks, oftentimes on a single time period using comparatively narrow datasets. We address this through three primary contributions. First, we introduce Market-1T, a dataset containing nearly one trillion observations across U.S. equities from 2008 to 2025 at 1 Hz resolution. Second, we develop and implement a rigorous evaluation protocol. Third, we conduct a systematic large-scale study of financial representation learning, comparing 18 encoder-training strategies across nearly two decades of market regimes. We evaluate learned representations both by their predictive utility on common finance tasks and through probes of latent structure. We find that encoders with similar predictive performance can organize market state very differently. Collectively, we establish a foundation for training and evaluating financial market representations in support of world models such as DINO-WM, V-JEPA 2, and LeWM.

• The paper introduces the Market-1T dataset, containing nearly one trillion one-second observations for U.S.-traded equities from 2008 through 2025, with 11 variables including bid/ask prices, transaction prices, and traded volume.
• The study evaluates 18 encoder-training strategies, finding that supervised multihead encoders excel in prediction but do not preserve firm, temporal, industry, or factor structure, highlighting the need for multi-objective representation
• The research shows a disconnect between predictive performance and latent-state organization in financial data, underscoring that model quality should consider multiple objectives like firm identity, day-level market conditions, industry relations, and factor exposure

Research objective and framing

“Towards Financial World Modeling” (2610.09048) treats financial world modeling as a representation-learning problem. A useful market state must support more than a single forecasting objective: it may need to encode expected returns, volatility, liquidity, firm-specific characteristics, market-wide conditions, and cross-asset dependence. The paper therefore asks whether encoder quality can be characterized by predictive performance alone, or whether forecasting and latent-state organization constitute distinct dimensions of representation quality.

The paper makes three contributions. First, it introduces Market-1T, a publicly available dataset containing nearly one trillion one-second observations for U.S.-traded equities from 2008 through 2025. Second, it develops an evaluation protocol designed to reduce regime-selection bias, probe-sample bias, and confounding from portfolio construction. Third, it compares 18 encoder-training strategies, including supervised models, self-supervised learning (SSL) methods, time-series foundation models (TSFMs), classical baselines, and an untrained ViT.

The central empirical claim is deliberately multidimensional: representations with similar forecasting performance can encode materially different market structures, while representations that preserve economically meaningful structure need not be the strongest forecasters. This distinction is directly relevant to world models, in which a state representation must support downstream dynamics and planning rather than only a fixed supervised prediction task.

Market-1T and the data representation

Market-1T is constructed from consolidated Securities Information Processor feeds, including CTA and UTP NBBO updates and consolidated trade reports. It covers the changing universe of U.S.-traded equities, including foreign firms listed on U.S. exchanges. The data begin in 2008, following the adoption of Regulation National Market System, and extend through 2025.

Each one-second observation contains 11 variables:

• four quote-state variables: bid price, bid size, ask price, and ask size;
• five transaction-price variables: open, high, low, close, and VWAP;
• two activity variables: traded volume and trade count.

Asynchronous quote and trade events are synchronized into one-second intervals. Quote variables use last-observation-carried-forward semantics, whereas transaction variables aggregate eligible trades within the interval. The authors apply financial aggregation rules when constructing lower-frequency views: ending quote states are retained, OHLC values are computed from eligible transactions, VWAP is recomputed from all constituent trades, and volume and trade counts are summed. This design allows temporal-scale augmentations without interpolation, preserving the economic meaning of the inputs.

The benchmark itself uses a liquid, dynamically recomputed universe. Securities must be common stocks, have prior-month average VWAP above $5, average daily dollar volume of at least $20 million, and trade during more than 20% of regular-trading-hours seconds. Because eligibility is determined using only preceding-month information, the benchmark avoids fixed-universe survivorship and look-ahead bias.

The dataset’s principal methodological value is not only its size. Its combination of 1 Hz resolution, broad cross-sectional coverage, and an 18-year span permits repeated evaluation across heterogeneous market regimes. This is important because the paper finds that regime variation dominates much of the apparent experimental variance.

Evaluation protocol

The authors evaluate models on 32 months sampled throughout 2008–2025. For each evaluation month, training uses only the preceding six months, with five separate months reserved for hyperparameter selection. TSFMs are an explicit exception to the six-month training restriction because their pretraining data and procedures are external to the benchmark.

The predictive evaluation probes the encoder representation with ridge regression and measures rank information coefficient (IC) for three 15-minute-ahead targets:

1. cross-sectional return;
2. volatility change;
3. bid–ask spread change.

Rank IC is chosen instead of Sharpe ratio as the primary predictive metric because it measures forecast ordering before portfolio construction and execution assumptions are imposed. The authors nevertheless evaluate 2,448 portfolio implementations spanning universe selection, portfolio weighting, risk modeling, rebalancing, and execution quality. They report that portfolio-design choices affect realized Sharpe substantially more than encoder choice, while IC remains positively related to Sharpe after averaging over the implementation harness.

Probe sample size is itself a major source of bias. Probe performance rises substantially with the number of fitting samples, implying that small probe datasets can underestimate information retained by an encoder. The paper therefore uses the full probe-training pool for headline results.

The evaluation also explicitly separates month-to-month variation from random-seed variation. Month explains 97.2% of variation for return prediction, 98.9% for volatility change, and 99.9% for spread change in the reported variance decomposition. Consequently, repeated runs within one market period provide a substantially less informative estimate of generalization than evaluation over many periods.

Figure 1: Rank IC changes substantially across market regimes, motivating evaluation over 32 sampled months rather than a narrow temporal slice.

This result has a direct implication for financial representation-learning benchmarks: single-period evaluations can produce systematically biased rankings of encoders. The authors reduce this problem by subtracting the monthly Random ViT floor when estimating standard errors, thereby isolating performance relative to the difficulty of each evaluation month.

Predictive representations

The headline forecasting results show a clear advantage for supervised multi-task training. The supervised Multihead encoder obtains rank IC values of 0.0311 for return, 0.0941 for volatility change, and 0.2455 for spread change under the ridge-probe evaluation. Its own prediction heads produce similar values: 0.0317, 0.0952, and 0.2441, respectively. These results indicate that the representation itself, rather than only an unusually effective task-specific head, accounts for much of the performance.

The strongest self-supervised methods are substantially weaker on return and volatility. TimeMAE reaches 0.0199 on return, 0.0706 on volatility change, and 0.1408 on spread change. Time-warped LeJEPA reaches 0.0204, 0.0721, and 0.1691, respectively, while TS2Vec reaches 0.0205, 0.0719, and 0.1500. The untrained Random ViT obtains 0.0170, 0.0694, and 0.1692. Thus, several SSL methods improve only modestly over a random encoder, and some fail to outperform it.

The spread-change results are particularly important. All SSL variants perform below the Random ViT on this target, with the exception that some methods approach the random baseline. The supervised Multihead model also underperforms the spread-specific supervised encoder, which obtains 0.2517. The authors interpret this as evidence that liquidity prediction contains specialized information that can be diluted by joint optimization with return and volatility objectives.

Encoder	Return IC	Volatility-change IC	Spread-change IC
Supervised return	0.0272	0.0695	0.1305
Supervised volatility	0.0197	0.0883	0.1480
Supervised spread	0.0178	0.0762	0.2517
Supervised Multihead	0.0311	0.0941	0.2455
LeJEPA, time warping	0.0204	0.0721	0.1691
TS2Vec	0.0205	0.0719	0.1500
TimeMAE	0.0199	0.0706	0.1408
Random ViT	0.0170	0.0694	0.1692

The implication is not simply that supervised learning dominates SSL. Rather, the objective determines which information is retained. A representation optimized for a particular forecast may discard variables useful for other market-state descriptions, whereas SSL objectives may impose invariances that remove target-relevant microstructure information.

Figure 2: The paper’s overview connects augmentation-induced latent organization with forecasting performance and shows that smaller ViTs can be more compute-efficient than larger models.

Augmentation and induced invariances

The augmentation experiments demonstrate that SSL is not objective-independent. Different positive-pair constructions impose different invariances on market observations.

Same-stock views encourage invariance across temporal crops and resolutions of a single stock-day. The resulting representation tends to preserve firm identity, often through bid- and ask-size distributions, but it does not necessarily encode day-level market conditions. Cross-stock pairing at the same time instead encourages retention of market-wide movements while suppressing idiosyncratic firm information. Cross-stock same-industry pairing introduces a stronger prior that economically related firms should share latent structure.

Time warping is among the strongest SSL choices for predictive probes. With LeJEPA, time warping yields 0.0204 return IC and 0.0721 volatility-change IC, placing it among the best non-supervised methods. However, it remains weak on spread change. This pattern indicates that temporal-scale invariance can preserve information relevant to broad forecasting while discarding or weakening fine-grained liquidity signals.

Figure 3: LeJEPA augmentations produce distinct invariances, with same-stock views emphasizing firm identity and cross-stock views emphasizing shared market movements.

The result supports a broader methodological conclusion: in financial SSL, the augmentation defines an economic hypothesis about what should remain invariant. Generic image-style augmentations cannot be assumed to be appropriate because temporal alignment, firm identity, industry membership, and market-wide conditions have different economic meanings.

Scaling and temporal decay

At a fixed training-FLOP budget, smaller encoders consistently outperform larger ones across the three forecasting targets. In particular, ViT-Tiny outperforms ViT-Base even when the latter is trained with up to ten times the compute. This is a strong and somewhat contradictory result relative to conventional scaling intuitions: within the studied regime, increasing parameter count without adding proportionate useful statistical information reduces compute efficiency.

The authors caution that these models are small relative to production-scale forecasting systems, so the result should not be interpreted as a universal law of financial-model scaling. It does establish, however, that model scale is not an adequate proxy for representation quality under a fixed compute budget.

Figure 4: Smaller ViT encoders outperform larger encoders across tasks at matched compute budgets, including comparisons in which ViT-Base receives substantially more training FLOPs.

Temporal decay is estimated by comparing a model trained earlier with an otherwise matched model trained closer to the evaluation date. This controls for changes in the intrinsic difficulty of the forecasting problem. Return representations are relatively persistent, whereas volatility-change and spread-change representations decay with temporal distance. Over five years, the reported degradation is approximately 20% of base IC for volatility change and approximately 10% for spread change.

Figure 5: Relative representation decay is small but detectable, especially for volatility-change prediction and, to a lesser extent, spread-change prediction.

The implication is that retraining requirements are target-dependent. A return representation may remain useful across longer intervals, while liquidity- and volatility-related representations are more sensitive to changes in market microstructure and regime.

Latent organization beyond prediction

The paper’s second evaluation axis examines whether distances in representation space reflect economically meaningful relationships. The authors test firm identity, shared trading day, industry relations, and statistical-factor exposure. They also measure effective rank using RankMe (Garrido et al., 2022), which quantifies how many directions carry substantial representation variance.

Supervised models have low effective rank. The single-task return, volatility, and spread encoders have RankMe values of 2.8, 3.0, and 3.5, respectively, while the Multihead encoder has a value of 3.7. Several SSL models occupy substantially higher-dimensional spaces: CoST reaches 17.3, TimeMAE 61.7, and TF-C 81.4. TSFMs are higher still, with Chronos-2 at 107.1 and TimesFM 3.0 at 111.9, although representation width differs across models and RankMe is descriptive rather than a direct quality score.

The organization tests reveal a trade-off between firm-level and day-level structure. Same-stock augmentations induce representations that cluster observations by firm, whereas cross-stock augmentations produce stronger day-level organization. The paper reports a strong negative relationship between own-firm and own-day centroid performance, indicating that an encoder generally cannot maximize both forms of invariance under the same training configuration.

The statistical-factor analysis provides a complementary test. Instead of defining relatedness using labels such as industry or day, it estimates factors from realized return comovement and asks whether firm factor loadings can be decoded from embeddings and whether embedding principal directions span the factor subspace. This evaluates whether the representation preserves cross-asset dependence relevant to covariance modeling.

Figure 6: Encoder rankings for forecasting and latent organization are weakly related, with Spearman correlation $\rho=-0.19$.

The overall correlation between forecasting rankings and organization rankings is $\rho=-0.19$, indicating that the two evaluation families are effectively unrelated. The supervised Multihead encoder is the only trained model that exceeds the Random ViT on every latent-space evaluation. Among the pretrained models, TimesFM 3.0 also exceeds the random baseline on the full latent suite.

This is the paper’s strongest conceptual result: forecasting quality is not a sufficient statistic for the usefulness of a financial state representation. A model can preserve information predictive of near-term returns while failing to organize observations according to firm, day, industry, or factor structure. For a world model intended to simulate consequences of actions, such omissions may be consequential because execution and risk depend on liquidity, covariance, and market-wide state in addition to expected returns.

Portfolio implementation and the choice of metric

The authors’ preference for IC over Sharpe is supported by a large implementation-harness experiment. Across 2,448 portfolio configurations, execution quality, rebalancing, universe selection, weighting, and risk modeling all materially affect Sharpe. At 100% effective-to-quoted spread, the range induced by rebalancing is 24.7 Sharpe points across the evaluated methods, compared with 13.3 for security selection and 11.9 for universe screening. The corresponding model-level variation is much smaller.

IC and Sharpe remain positively associated, but the relationship weakens as execution costs increase. The cross-method correlation between IC and average Sharpe is 0.92 at zero effective spread and 0.46 at 100% effective spread. Within individual portfolio implementations, the average correlation is lower, falling from 0.52 to 0.29 over the same range.

These results imply that Sharpe is appropriate as a downstream robustness check but unstable as the primary representation benchmark. A representation comparison based on a single portfolio specification would conflate encoder quality with assumptions about execution, turnover, risk, and portfolio construction.

Figure 7: Realized Sharpe depends strongly on effective execution costs and the portfolio-construction harness, supporting IC as the primary representation metric.

Limitations and open questions

The study evaluates a substantial but specific market domain: consolidated U.S. equity quotes and trades, primarily within a liquid common-stock universe. The benchmark excludes many instruments and market settings, including options, futures, fixed income, cryptocurrencies, less liquid equities, and detailed order-level event histories. Consequently, the findings do not establish that the same objective–augmentation relationships hold outside U.S. equities or at event-level resolution.

The dataset also represents the market through 11 aggregate features. Although these preserve one-second NBBO and transaction information, they omit participant identities, order-book depth beyond the best quote, venue-level routing, cancellations, queue position, and explicit actions. The proposed encoders are therefore state representations for observed market aggregates, not complete market states in the microstructural sense.

The world-modeling claim remains preparatory. The paper evaluates encoder representations but does not train a full action-conditioned dynamics model, test counterfactual execution, or demonstrate planning under interventions. Whether a representation that performs well on the latent organization suite supports accurate dynamics prediction and policy optimization remains open.

Comparisons to TSFMs are also not fully symmetric. These models are exempt from the six-month training restriction, have different widths, and in some analyses are evaluated at layers selected through sweeps that are not applied uniformly to all encoder families. The authors acknowledge this asymmetry. Similarly, the strong conclusion about model scale is limited to the tested architecture, training schedule, and compute range.

Finally, probe-based evaluation measures information accessible to a specified ridge readout, not all information theoretically present in an embedding. Conversely, strong probe performance does not establish causal sufficiency for trading or planning. The paper’s conclusions are therefore best interpreted as benchmark-relative claims about accessible predictive and organizational structure.

Conclusion

“Towards Financial World Modeling” (2610.09048) establishes a large-scale empirical foundation for studying financial state representations. Market-1T provides unusually broad temporal and cross-sectional coverage at one-second resolution, while the evaluation protocol demonstrates that regime diversity, probe sample size, and implementation assumptions materially affect conclusions.

The main findings are that supervised multihead encoders provide the strongest forecasting representations, smaller ViTs can be more compute-efficient than larger models, and representation decay is target-dependent. More importantly, forecasting and latent organization are distinct: their method rankings have correlation $\rho=-0.19$, and successful prediction does not guarantee preservation of firm, temporal, industry, or factor structure. The paper therefore frames financial world modeling as a multi-objective representation problem in which state quality must be assessed by the market information and relational structure it preserves, not by return prediction alone.