| Current deep learning models for limit order book (LOB) prediction treat every time step as equally predictable, ignoring the microstructure regime dynamics that govern order flow. We develop a cross-attention Transformer that fuses LOB state with Hawkes intensity features, evaluated through a 2×2 factorial ablation varying fusion architecture (concat vs. cross-attention) and auxiliary content (handcrafted vs. Hawkes) independently on 18 days of NVIDIA tick data. The decomposition attributes the dominant share of accuracy improvement over an LOB-only baseline to cross-attention fusion (+6.97 pp on average), with auxiliary content contributing +0.82 pp; the full Hawkes-XAttn system achieves +7.38 pp (p < 0.0001, Cohen’s d = 7.86). Within crossattention, Hawkes content adds incremental accuracy (+1.26 pp, p= 0.027),sharpeningoftheno-tradesignalduringintensitybursts,andpartial mitigation of an architectural latency penalty. Cross-attention configurations learn regime awareness internally—maintaining or improving accuracy during bursts where concat-fusion configurations degrade—and our controlled latency analysis shows they retain their accuracy advantage at every latency level tested despite the architectural latency penalty. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.