Estrada Maldonado, J. O.
EN ES
← All research

A two-tower early fusion architecture for Liga MX football outcome prediction

Combining transfer-learned embeddings and tabular statistics, with a Club América case study

Jesús Osvaldo Estrada Maldonado — sole author

Forthcoming · presented December 2026 ICECER2026 · Paper 557 Scopus indexing expected January 2027 English
This page is published ahead of presentation so the work can be found and cited. Figures are final as submitted; the camera-ready reference will be added once proceedings are issued.

The problem

Automated football outcome prediction has run into a performance ceiling that two decades of methodological refinement have not displaced. The most competitive models sit inside a narrow accuracy band that purely tabular methods have never surpassed, and deep networks trained from scratch usually make things worse: with a small sample, they overfit.

The architecture

The paper proposes and theoretically grounds a hybrid classifier built on a two-tower early fusion design.

  • Tabular tower. Historical match statistics encoded by a multilayer perceptron trained end to end on eight complete seasons of the Mexican first division.
  • Text tower. Pre-match media coverage encoded by a large-scale pre-trained language model specialised in Spanish social media text.
  • Fusion. Each network is truncated after its last hidden layer; the two resulting dense embeddings are concatenated directly into a single feature vector.
  • Head. The fused vector feeds a gradient boosting classifier.

The reason the design is expected to work is transfer, not capacity: the text encoder's linguistic knowledge is imported rather than learned from a small sports corpus, which is what mitigates the overfitting that sinks from-scratch deep models on data of this size.

Match statistics 8 seasons, Liga MX Pre-match media text Spanish social coverage MLP tower trained end to end Pre-trained LM transfer-learned, reduced Concatenation single fused vector Gradient boosting win / draw / loss
Early fusion: both towers are truncated after their last hidden layer and their embeddings joined before classification, rather than combining predictions after the fact.

Case study and evaluation

The case study covers 350 fixtures of Club América, selected because the club generates the greatest media discussion in Liga MX and therefore carries the richest pre-match textual signal. Results are averaged over 100 independent training runs and validated by an independent cross-validation.

Results

Tabular tower only52.9%
Two-tower fusion57.1%
Accuracy, mean of 100 independent runs. The gain is statistically significant at p < 0.0001, and precision, recall, F1 and log loss all improve simultaneously.

The improvement holds once the text embedding is dimensionality-reduced to fit the available sample size — without that reduction, the extra width costs more than the signal is worth. Draw recall, notably, remains essentially unimproved: the fused representation helps separate wins from losses, not the ambiguous middle.

The methodological finding

The evaluation protocol turned out to matter as much as the architecture. A single training run was not statistically distinguishable from no effect at all. Only the multi-run protocol revealed the improvement. On football datasets of comparable size, rigorous repeated-run evaluation is not a nicety — it is a precondition for knowing whether anything happened.

If a result on a few hundred fixtures is reported from one training run, the honest reading is that it has not been measured.

Why it generalises

Nothing in the design is specific to football. The pattern — a small tabular dataset, an adjacent stream of unstructured text, and a pre-trained encoder that imports linguistic knowledge the local corpus cannot supply — recurs across applied prediction problems in economics and the social sciences, which is where I came to it from.