This paper introduces SimLOB, a novel Transformer-based autoencoder designed to learn compact vectorized representations of Limit Order Book (LOB) data for financial market simulation (FMS) calibration. Unlike prior work that only uses mid-price data, SimLOB captures the full market microstructure by encoding LOB's high-dimensional, multi-scale, and precedence-constrained structure into a low-dimensional latent vector. The encoder consists of a fully connected network for feature extraction, a Transformer stack to model non-linear temporal dependencies, and another fully connected network for dimension reduction; the decoder is symmetric. The model is trained to minimize reconstruction error on LOB segments of 100 time steps.

Extensive experiments on synthetic data from the PGPS agent-based model and real stock data demonstrate that SimLOB significantly outperforms existing LOB-oriented neural networks (e.g., CNN+LSTM) in reconstruction quality. Crucially, when the learned latent representations are used as the calibration objective for FMS, they yield higher-fidelity simulations compared to traditional mid-price calibration or raw LOB inputs. Ablation studies confirm the importance of the Transformer and fully connected layers over convolutional alternatives. This work represents the first successful attempt to calibrate FMS using LOB data, enabling more accurate market simulations for anomaly detection and policy analysis.