Forma transformer outperforms frontier LLMs on 20-quarter financial forecasts

Most firm value in discounted-cash-flow valuation sits beyond one year, yet no prior system jointly forecasts complete financial statements past that window. A new paper from Travis L. Johnson and five collaborators introduces Forma, a transformer that reads financial statements as sets of (account, quarter, value) tuples and maximizes a masked-tuple Gaussian likelihood. The model predicts 78 statement line items across 1-20 quarters ahead, relying only on past statements and an industry code. The authors also release ProForma-20Q, a reproducible benchmark scored by change-space R², with model weights available on arXiv.

Forma beats every competitor fielded: classical machine learning, chained gradient boosting, a zero-shot time-series foundation model, and frontier large language models. Notably, its advantage grows with forecast horizon—exactly where valuation accuracy matters most. The model's Gaussian predictive intervals never under-cover, and its forecasts nearly satisfy accounting identities; exact coherence can be restored with no statistically significant accuracy loss. Because Forma uses a tuple interface, it supports scenario analysis without retraining—pinning future revenue paths sharpens forecasts across the rest of the statement. This combination of long-horizon accuracy, reliable uncertainty, and flexible scenario testing makes Forma a significant step for AI-driven financial analysis.