Systematic review · scoped evidence map
How should exogenous variables enter a transformer forecast—and what would it take to trust that design in finance? This concise review maps the core method, adjacent financial evidence, and the validation gaps that remain.
Anchor papers
TimeXer isolates the endogenous target from exogenous inputs instead of treating every variable symmetrically. Its design introduces dedicated embeddings and attention pathways for the practical exogenous-forecasting setting.
A domain-specific application that tests transformer-style forecasting on market data. It is useful as adjacent evidence, but it does not by itself establish that a TimeXer architecture will generalize across assets, regimes, or trading costs.
The survey organizes explainability approaches used in financial forecasting and exposes a key deployment requirement: predictive gains are insufficient without stable, decision-relevant explanations.
Cross-reference
| Question | TimeXer evidence | Finance requirement | Next study |
|---|---|---|---|
| How are outside signals represented? | Dedicated exogenous pathway and embedding design. | Respect release timing and revisions. | Point-in-time data ablation. |
| Does accuracy transfer to value? | Forecasting benchmarks support predictive evaluation. | Returns after turnover, slippage, and risk limits. | Walk-forward economic backtest. |
| Can users trust the signal? | Attention structure offers inspection points. | Stable explanations under regime change. | Explanation stability benchmark. |
Gaps
Provenance
Scope note: this is a curated anchor review, not the 40+ paper / 100k-word output described in the source post.