发表机构
University of Texas at Austin（德克萨斯大学奥斯汀分校）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究发布了ProForma-20Q基准，提出Transformer模型Forma，在1至20个季度的完整财务报表预测任务中，击败各类对比模型，优势随期限扩大，且能支持无需重训的情景分析。
AI 中文摘要
在财务报表预测任务中，专家式训练优于通用模型的规模扩展。据我们所知，目前尚无研究能联合预测超过一年的完整财务报表，而在折现现金流估值中，企业的大部分价值都来自这一期限之后的区间。我们发布了ProForma-20Q，这是一个可复现的基准，用于预测匿名企业在1至20个季度后的78项报表科目，输入为历史报表和行业代码，评估指标为变化空间R²。在该基准上，Forma（一种将报表读取为（科目、季度、数值）元组集合并最大化掩码元组高斯似然的Transformer模型）击败了我们参与对比的所有竞争对手：经典机器学习方法、链式梯度提升模型、零样本时间序列基础模型以及前沿大语言模型。其优势随预测期限延长而扩大，而估值最需要这种准确性，且其高斯预测区间从未出现覆盖不足的情况。Forma的预测几乎满足会计恒等式；实现完全一致性不会带来统计上显著的准确性损失。其元组接口支持无需重新训练的情景分析，我们还证明，固定未来收入路径可优化报表其余部分的预测。
英文摘要
Specialist training beats generalist scale when forecasting financial statements. To our knowledge, no prior work jointly forecasts complete financial statements beyond one year, yet in a discounted-cash-flow valuation most firm value sits past that window. We release ProForma-20Q, a reproducible benchmark for forecasting 78 statement line items 1-20 quarters ahead, for anonymized firms, from past statements and an industry code, scored by change-space $R^2$. On it, Forma, a transformer that reads statements as sets of (account, quarter, value) tuples and maximizes a masked-tuple Gaussian likelihood, beats every competitor we field: classical machine learning, chained gradient boosting, a zero-shot time-series foundation model, and frontier large language models. Its lead widens with horizon, where valuation needs accuracy most, and its Gaussian predictive intervals never under-cover. Forma's forecasts nearly satisfy accounting identities; exact coherence is recoverable at no statistically significant accuracy cost. Its tuple interface supports scenario analysis without retraining, and we show that pinning future revenue paths sharpens the rest of the statement.
Comments46 pages, 2 figures, 3 tables. Benchmark: https://github.com/forma-lab-mccombs/proforma-20q. Model and weights: https://github.com/forma-lab-mccombs/forma-release
