Reconstruction Buys Variance, Not Independence: A Tokenizer Keeps the Microstructure That Duplicates OHLCV and Drops the Signed Channels

Abstract

Two-stage models for financial time series fit a tokenizer under a reconstruction objective and then an autoregressive model over the resulting codes. This assumes that a representation good enough to reconstruct from is good enough to forecast from, and it fails in a predictable direction. Reconstruction allocates code capacity by variance and covariance and never by downstream value, so a low-variance channel weakly covariant with the channels already coded is excluded deterministically rather than by chance. That is the statistical profile of trade-flow microstructure against price. On an entropy-calibrated fixture, a planted signal of known information content is lost to tokenization while the same quantity planted directly in token space is recovered by the identical backbone, localizing the loss to the tokenizer–backbone interface, not to model capacity. Re-specifying that interface restores per-bar legibility at unchanged bits per token, and we convert that property into a pre-registered hard stop on real data. The gate fired, and it named the channels. Almost all of the shortfall falls on the two signed channels, trade-flow imbalance and signed count imbalance, while the magnitude channels, which co-vary with volume, essentially clear it. The tokenizer preserves the microstructure that duplicates what OHLCV already carries and loses the microstructure that is independent of it, which is the part a microstructure model would want. That split was not anticipated, and it is the paper's result. The firing was pre-registered as the mechanism's own modal prediction, and it was obeyed: the five-cell trading ablation it gates was never run, and this paper reports no economic outcome. The finding is bounded by one budget point, one objective and one architecture: it establishes that this tokenizer does not carry the signed channels, not that none can.