Neural Sabermetrics World Model
- The paper demonstrates a unified world model that predicts pitch types and swing decisions from serialized baseball tracking data.
- The method treats baseball games as long auto-regressive sequences by converting events into language tokens for holistic simulation.
- It leverages continuous pretraining on over 10 years of MLB data, achieving competitive accuracy and outperforming traditional baselines.
Searching arXiv for the specified paper to ground the article in the cited source. Neural Sabermetrics with World Model denotes a LLM based play-by-play world model for baseball that reinterprets baseball tracking data as a language and trains a generative model of how baseball games unfold pitch by pitch (Ahn et al., 2 Feb 2026). In this formulation, baseball games are cast as long auto-regressive sequences of events, allowing a single model to condition on the full preceding context and predict multiple aspects of game evolution within a unified framework. The approach is positioned against both classical sabermetrics, which is described as powerful for retrospective analysis and player valuation but fundamentally descriptive rather than generative, and prior baseball machine learning models that remain narrow, predict isolated targets, and rely on engineered features and shallow temporal context (Ahn et al., 2 Feb 2026).
1. Conceptual reframing of sabermetrics
Classical sabermetrics has profoundly shaped baseball analytics by summarizing long histories of play into compact statistics. In the world-model formulation, those metrics remain valuable for valuation and retrospective analysis, but they do not define a generative model of how baseball games unfold pitch by pitch (Ahn et al., 2 Feb 2026). The central claim is therefore not that descriptive metrics are displaced, but that they are insufficient for simulating or predicting the next pitch, swing decision, or plate-appearance outcome.
The paper’s central idea is to reinterpret baseball tracking data as a language and train a LLM as a play-by-play world model (Ahn et al., 2 Feb 2026). Rather than treating each pitch or decision as an isolated supervised example, the model learns from full sequential context. This shifts the analytic emphasis from post-hoc summary statistics and single-step prediction toward a unified predictive model of game dynamics that can handle long context, multiple prediction tasks, and distribution shifts such as the regular season → postseason transition.
This suggests a conceptual continuity with world-model perspectives in reinforcement learning and control: the baseball environment is represented as an evolving sequence whose future states are conditioned on prior states and actions. A plausible implication is that sabermetric analysis becomes not only descriptive and evaluative but also simulation-oriented when framed through autoregressive sequence modeling.
2. Event serialization and sequence representation
The game is serialized into a long textual sequence of baseball events (Ahn et al., 2 Feb 2026). Each game contains global context, including teams, venue, and weather; inning/game state, including outs, score, and baserunners; pitch-level information, including pitch type, release characteristics, location, and outcome; and batter/pitcher decisions such as swing or take. Instead of tabular rows, the model receives chronologically ordered text tokens describing the full game state and pitch history.
The paper explicitly frames this as modeling baseball as a long auto-regressive sequence of discrete and continuous events, where each next event is conditioned on all previous events (Ahn et al., 2 Feb 2026). That representation is central to the claim that sequential sports data can be treated like language. The model therefore operates over the full event stream rather than over a manually curated subset of features for each task.
This serialized representation also explains the breadth of the downstream prediction space. Although the main evaluated tasks are pitch type prediction and batter swing decision prediction, the event language also contains pitch outcomes, game state, and other event information. The framework is therefore described as naturally supporting prediction of next pitch, swing decisions, plate-appearance evolution, pitch outcomes, and potentially richer outcomes in future work (Ahn et al., 2 Feb 2026).
3. Continuous pretraining, scale, and objective
The model is continuously pretrained on MLB tracking data collected from Statcast and MLB StatsAPI (Ahn et al., 2 Feb 2026). Reported scale includes more than 10 years of MLB tracking data, over 7 million pitch sequences, and approximately 3 billion tokens. The training uses 11 years of regular-season data, while postseason games are held out for evaluation.
Because full games are very long, averaging over 120K tokens per game, the authors split them into sliding windows of 3072 tokens without overlap (Ahn et al., 2 Feb 2026). The backbone is Llama-3.2 3B, with batch size 1024 and sequence length 3072, trained on TPU-v4-64. The paper does not introduce a new architecture from scratch; it uses a pretrained Llama-family model and continues pretraining it on serialized baseball sequences.
The training objective is standard next-token prediction, or autoregressive maximum likelihood. Formally, the model learns
and is trained to maximize
Downstream prediction is described as conditional generation or constrained decoding: given the context up to a pitch, the model autoregressively generates the next event token and maps it back to a structured label (Ahn et al., 2 Feb 2026). This is significant because the same likelihood-based backbone serves multiple predictive tasks without task-specific architectural redesign.
4. Tasks, evaluation regime, and baselines
The main evaluated tasks are pitch type prediction and batter swing decision prediction (Ahn et al., 2 Feb 2026). Pitch type prediction targets the next pitch category, such as four-seam fastball, slider, or changeup. Batter swing decision prediction targets whether the batter will swing or take. The broader framework, however, is presented as more general than these two tasks because the serialized representation includes richer event information.
A key feature of the evaluation protocol is the distinction between in-distribution and out-of-distribution testing. The model is trained on regular-season data only and evaluated on postseason games, which is explicitly treated as an out-of-distribution test because postseason baseball differs strategically and contextually from regular-season play (Ahn et al., 2 Feb 2026). The paper emphasizes that prior work typically treats postseason as harder and distribution-shifted, so performance under this split is used to support the claim that the world-model formulation generalizes.
The baselines are strong neural prior-work models reproduced under the same data span for fair comparison (Ahn et al., 2 Feb 2026). They consist of an RNN/LSTM baseline from Pi (2018) for pitch prediction, an attention-based LSTM from Yu et al. (2022) for multi-class pitch prediction, and an MLP baseline from Gopal et al. (2024) for swing/decision prediction. The paper’s positioning is therefore comparative rather than purely demonstrative: a single backbone model is evaluated against specialized neural baselines under matched temporal coverage.
5. Quantitative performance and empirical analyses
The paper reports that the resulting model is capable of predicting multiple aspects of game evolution within a unified framework and outperforms the performance of existing baselines despite using a single backbone model (Ahn et al., 2 Feb 2026). The abstract states that the model correctly predicts approximately 64% of next pitches within a plate appearance and 78% of batter swing decisions.
For pitch type prediction, the reported comparison is as follows:
| Model | Accuracy | Recall | F1 |
|---|---|---|---|
| Baseline | 0.633 | 0.792 | 0.720 |
| Ours | 0.637 | 0.792 | 0.722 |
The improvement is slight but favorable, especially in F1 (Ahn et al., 2 Feb 2026). For batter swing decision prediction, the reported in-zone and out-of-zone accuracies are:
| Model | IZ Acc | OZ Acc |
|---|---|---|
| Baseline | 0.325 | 0.704 |
| Ours | 0.766 | 0.792 |
The paper also reports 83.8% of at-bats contain at least one correct prediction and 54.7% contain at least two correct predictions (Ahn et al., 2 Feb 2026). This is used to show that correct predictions often persist across multiple pitches in the same at-bat, although accuracy drops as the number of correct predictions required in an at-bat increases.
Several analyses are presented to explain why the world-model formulation helps. Error accumulation over long sequences indicates that long-horizon autoregressive modeling is harder, but still captures useful sequential dependencies. Pitcher repertoire complexity matters: accuracy is highest for pitchers with smaller arsenals, reaching up to 66.8% accuracy for pitchers with two to three pitch types, and performance degrades as the number of pitch types increases (Ahn et al., 2 Feb 2026). Confusion matrix analysis shows bias toward four-seam fastballs, the most common pitch type, and frequent confusions between slider and sweeper and between sinker and changeup. Error distribution analysis further notes that four-seam fastballs and sliders dominate total errors mainly because they are frequent, while rare pitch types produce fewer errors, partly due to lower sample counts and more distinctive signatures.
Together, these analyses support the claim that the LLM is learning a game-level probabilistic structure, not just a static classifier (Ahn et al., 2 Feb 2026). This suggests that the model internalizes both frequency priors and physically or strategically related pitch structure.
6. Significance, applications, and limitations
The broader significance is framed as a step toward predictive, generative sports analytics (Ahn et al., 2 Feb 2026). The method moves beyond descriptive sabermetrics, provides a unified framework for multiple baseball prediction tasks, demonstrates that LLMs can act as world models for structured sports data, and suggests that sequential sports data can be treated like language.
Potential uses listed in the paper include pitch sequencing analysis, swing/approach prediction, at-bat simulation, opponent strategy modeling, postseason robustness studies, extensible multimodal baseball forecasting, and eventually run expectancy and batted-ball trajectory modeling (Ahn et al., 2 Feb 2026). These are presented as possible applications rather than fully realized outputs of the current evaluation protocol.
The limitations are explicit. Context-length limits force long games to be split into fixed windows, breaking some long-range dependencies. It remains open whether tokenized event representations or latent state abstractions are better. Evaluation is currently limited mostly to pitch and swing prediction. Richer outcomes like batted-ball trajectories and run expectancy are not yet modeled (Ahn et al., 2 Feb 2026). These constraints qualify the scope of the contribution: the work demonstrates a viable play-by-play world model, but not a complete generative account of all relevant baseball outcomes.
A common misconception would be to treat the framework as merely a relabeling of standard sequence classification. The paper argues otherwise by grounding prediction in serialized full-game context and by emphasizing conditional generation over long auto-regressive event streams (Ahn et al., 2 Feb 2026). Another misconception would be to assume that classical sabermetrics is rejected; the stated position is narrower, namely that classical summaries are invaluable but do not themselves constitute a generative process.
7. Position within baseball modeling
The paper’s contribution is to show that MLB tracking data can be serialized into a long event language and used to continuously pretrain a Llama-based model as a baseball world model (Ahn et al., 2 Feb 2026). In contrast to prior systems that predict pitch type, swing/no-swing, or pitch outcome in isolation, the world-model perspective consolidates multiple tasks under one autoregressive backbone trained with standard likelihood objectives.
Within baseball modeling, the significance lies less in architectural novelty than in representational unification. A pretrained Llama-family model is adapted to baseball by continuous pretraining on serialized event streams rather than by constructing a bespoke sports architecture from scratch (Ahn et al., 2 Feb 2026). This supports the view that foundation-model techniques can be transferred to structured sports data when the domain is expressed as a sufficiently rich token sequence.
The paper’s bottom line is therefore specific: with standard autoregressive likelihood training, the model predicts next pitch type and swing decisions well, generalizes to postseason games, and provides a single backbone for multiple baseball prediction tasks (Ahn et al., 2 Feb 2026). A plausible implication is that future research may extend this paradigm from pitch-by-pitch forecasting toward fuller simulation-oriented sports modeling, but the demonstrated claim is confined to the play-by-play world model and its reported predictive performance.
