Long-Horizon Forecasting of Complete Financial Statements with Forma

Battle Count: 50

Highly relevant for fundamental quantitative strategies that rely on pro forma financial statement forecasts for DCF valuation, earnings prediction, and credit risk assessment. The paper demonstrates that a small specialist model (~0.9M parameters) significantly outperforms frontier LLMs and generalist foundation models for this task, with the advantage widening at longer horizons where most firm value sits. The scenario analysis capability (pinning revenue paths) is directly applicable to analyst workflows. However, the paper explicitly excludes stock returns and analyst forecasts as features, and focuses on quarterly U.S. filings, limiting direct applicability to high-frequency trading. The change-space R² metric and probabilistic outputs are valuable for risk-adjusted position sizing. The benchmark (ProForma-20Q) provides a standardized evaluation framework for fundamental forecasting models used in quantitative investing.

Moderate complexity. The Forma architecture itself is relatively simple: a 4-layer Transformer encoder with d_model=128 and ~0.9M parameters, trained for 12 epochs on a single A100 GPU (~16.5 hours per seed, ~$120 total for 5 seeds). The tuple-set representation and masking procedures (identity-aware grouped masking, pinned-future masking, ex-ante future grid) require careful implementation. The data pipeline (WRDS extraction, YTD-to-quarterly conversion, asinh standardization, temporal splits with purged targets) is well-documented and reproducible with one command. The main complexity lies in the data preprocessing and evaluation protocol rather than the model architecture. Training the 1,560 per-item-per-horizon models for tabular baselines (elastic net, RF) is computationally intensive (~40 hours on one GPU for RF).

Exceptional reproducibility: full pipeline, configurations, evaluation code, and documentation released on GitHub (proforma-20q and forma-release repos). Data environment rebuildable from WRDS with one command, verified by published checksums. Trained weights and seeded regeneration scripts for all learned competitors provided. Exact prompts for LLM benchmark included. All hyperparameters documented in Table 7. Temporal splits with purged targets ensure no look-ahead leakage.