Researchers have proposed a Transformer model that adjusts how it weighs financial signals depending on what the market is doing.
The Adaptive Financial Transformer, or AFT, encodes 95 engineered financial features into 11 semantic categories — things like momentum, volatility, and volume — then uses a gating mechanism to shift attention based on inferred market regimes. The idea is that what mattered in a trending market is different from what matters in a choppy one, and a static model can't capture that. The paper also flags a methodological problem common in this research space: sequence alignment and backtesting errors that make trading strategies look better than they are, and claims to correct for both.
Most financial Transformer research benchmarks against models that treat all inputs equally, which is a known weakness when markets shift behavior. AFT's regime-aware attention is a structural attempt to fix that, and the 15.2% reduction in model complexity is a practical argument for deployment — smaller models are cheaper to run and easier to audit. The paper also introduces a composite training objective that jointly targets prediction error, directional accuracy, and a non-overlapping Sharpe ratio, which is a more honest optimization target than raw return.
Competitive results against recurrent networks and vanilla Transformers are encouraging, but the financial ML graveyard is full of models that backtested well and collapsed in production — the authors' own caveat about backtesting bias is a useful reminder of exactly that.