CAME: Company-Aware Evidence-Memory Experts for Interpretable Quarter-Ahead Revenue Forecasting¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/yy1200/came-revenue-forecasting
Area: LLM Reasoning
Keywords: Complex Numerical Reasoning, Residual Forecasting, Evidence-Memory Experts, Revenue Prediction, Interpretable Forecasting
TL;DR¶
CAME reformulates quarter-ahead corporate revenue prediction as an anchor-relative log-residual refinement task over a no-leakage statistical baseline, combining typed semantic evidence cards, guidance-gated memory correction, and multi-channel experts to achieve superior numerical precision and source-linked auditability across 336 company-quarters.
Background & Motivation¶
In financial natural language processing, large language models are increasingly tasked with understanding complex corporate disclosures, earnings conference calls, and regulatory filings. However, while language models excel at intermediate linguistic tasks such as sentiment analysis, question answering, and qualitative summarization, predicting absolute fundamental business metrics—specifically total quarter-ahead corporate revenue—poses severe calibration and verification bottlenecks. Unconstrained end-to-end numerical generation by LLMs suffers from acute Calibration Failure: lacking rigid mathematical constraints tied to a company's unique revenue scale, minor phrasing nuances or retrieval variations can trigger multi-billion-dollar hallucinations. Furthermore, ungrounded direct generation exhibits Provenance Failure, leaving risk-averse financial stakeholders without any verifiable trail connecting specific textual disclosures to the emitted numerical prediction.
Conversely, traditional time-series statistical models and econometric forecasters operate with scale stability by fitting historical numerical trajectories. Nevertheless, these pure numerical methods suffer from Contextual Blindness. By relying strictly on lagged quarterly records, they remain blind to forecast-time operational dynamics that appear in narrative transcripts long before impacting lagging quarterly statements—such as product architecture transitions, supply chain bottlenecks, channel inventory destocking, and forward-looking executive guidance. When companies enter macroeconomic shifts or structural regime changes, history-only anchors become dangerously stale and directionally sluggish.
Bridging this gap requires dismantling the assumption that text models must invent the absolute revenue scale from scratch. Instead, quantitative scale stability and qualitative semantic awareness must be explicitly decoupled. Core idea: formalize quarter-ahead revenue prediction as an anchor-relative log-residual refinement task (\(\hat{\delta}_i = \log(y_i) - \log(a_i)\)) on top of a scale-aligned statistical prior, employing guidance-gated anchor-memory correction, typed evidence cards, and a multi-expert mixture to produce traceable, guardrailed numerical adjustments.
Method¶
Overall Architecture¶
CAME operates through four deterministic, strictly ordered stages: first, raw narrative disclosures are distilled into structured, typed Evidence Cards grounded in specific company reporting segments and causal revenue relation families; second, a no-leakage statistical time-series model establishes an initial prior, which is adjusted by a guarded prior-error memory mechanism in scenarios lacking management guidance to form the Guidance-Gated Base Anchor; third, a mixture of Evidence-Memory Experts proposes anchor-relative log-residuals alongside heuristic reliability weights; fourth, a reliability-weighted composition applies a Guidance Guardrail to attenuate revenue-space displacement according to disclosure uncertainty, outputting the final calibrated prediction alongside an end-to-end provenance trace.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Inputs: Disclosures, Historical Lags & Company Config"] --> B["Stage 1: Evidence-Card Distillation<br/>Extract causal drivers, segments & polarities"]
B --> C["Stage 2: Guidance-Gated Base Anchor<br/>No-leakage statistical prior + error memory"]
C --> D["Stage 3: Evidence-Memory Experts<br/>Explicit guidance, current evidence & temporal memory"]
D --> E["Stage 4: Revenue-Space Guardrails<br/>Uncertainty attenuation & provenance trace"]
E --> F["Outputs: Calibrated Revenue Forecast & Source-Linked Trace"]
Key Designs¶
1. Evidence-Card Distillation: Structured Semantic Abstraction of Narrative Disclosures
Targeting the core vulnerability of unconstrained generation to hallucinations and unverified numbers, CAME rejects passing raw transcript documents directly into the numerical forecasting module. Instead, an information extraction pipeline transforms observable text into typed semantic artifacts \(v = (x, h, \varsigma, r, p, s, \xi, \gamma, \rho) \in E_i\). Here, \(x\) represents the verbatim transcript span, \(h\) denotes the affected horizon, \(\varsigma\) is the normalized company segment (e.g., DCAI, CCG, or Intel Foundry), \(r\) specifies the revenue-mechanism relation family (e.g., demand, supply constraints, pricing, macro), \(p \in \{+, -, \pm, \text{unk}\}\) records revenue-effect polarity, \(s\) indicates qualitative strength, \(\xi\) provides a persistence hint, \(\gamma\) is an extraction confidence score, and \(\rho\) classifies the functional role as realized or forward-looking. Governed by declarative company schemas and deterministic validators, this extraction rejects generic optimism and pure cost commentary lacking revenue pathways, while enforcing strict temporal cutoff eligibility (\(r_j \le t_i\)) to prevent look-ahead bias.
2. Guidance-Gated Base Anchor: Systematic Historical Bias Rectification in No-Guidance Regimes
Addressing the severe instability and scale drift of statistical models when corporate management provides no numerical guidance, CAME introduces an Anchor-Memory Correction layer on top of a dynamically chosen statistical anchor \(b_i\) (selected online without leakage from an admissible family including ARIMA, ETS, LightGBM, and Random Forest). When target-quarter total revenue guidance is absent, the system inspects the signed chronological sequence of prior no-guidance anchor residuals. If these residuals satisfy a signed-strength admission gate (requiring \(\ge 6\) historical no-guidance quarters and an average magnitude exceeding threshold \(\tau_{\text{AM}} = 0.02\)), a bounded log-space correction \(\Delta_{\text{mem}, i} = G_{\text{mem}, i} \cdot m_i\) is applied; otherwise, the gate remains closed. The base anchor is defined as: $\(a_i = b_i \exp(\Delta_{\text{mem}, i})\)$ Crucially, \(G_{\text{mem}, i}\) is strictly zero whenever explicit or non-explicit management guidance is present, ensuring that base anchors retain their unaltered statistical properties whenever corporate management has already provided forward numerical bounds.
3. Evidence-Memory Experts: Multi-Channel Coordination of Text, Analogs, and Guidance
Targeting the challenge of reconciling heterogeneous textual drivers with quantitative guidance, conditioned on the base anchor \(a_i\), CAME deploys three specialized expert channels \(K = \{\text{cur}, \text{tmp}, \text{guid}\}\) that propose anchor-relative log-residuals \(d_i^k\) along with heuristic reliability scores \(\sigma_i^k \in [0, 1]\): - Explicit-Guidance Expert: Reconciles the statistical prior with management-provided numeric guidance midpoints \(g_{\text{mid}, i}\), proposing \(d_i^{\text{guid}} = \log(g_{\text{mid}, i} / a_i)\) with reliability governed by guidance range tightness and historical guidance forecasting accuracy. - Current-Evidence Expert: Evaluates the immediate quarter's evidence cards using a partitioned learning target—learning the entire residual relative to \(a_i\) for no-guidance rows, but only the marginal residual unexplained by guidance when guidance is present. Its reliability is dynamically scaled based on historical same-bucket win rates and marginal sMAPE improvements. - Temporal-Memory Expert: Overcomes keyword-matching fragility by executing typed cross-card semantic matching across segment tags, relation families, and polarity alignment against historical quarters. It retrieves the top \(K_q = 3\) historical analog quarters and averages their realized base-anchor residuals \(\epsilon_j^{\text{hist}} = \log(y_j) - \log(a_j)\) under temperature-scaled weighting. The channel proposals are aggregated via reliability-weighted pooling into a pre-guardrail residual \(\hat{\delta}_i = \sum_{k \in K} w_i^k d_i^k\).
4. Revenue-Space Guardrails: Bounded Conservatism and Transparent Traceability
To mitigate the risk of qualitative text triggering wildly uncalibrated numerical over-corrections, CAME routes the pre-guardrail estimate \(\hat{y}_i^{\text{pre}} = a_i \exp(\hat{\delta}_i)\) through a revenue-space attenuation function: $\(\hat{y}_i = a_i + \alpha(\chi_i^{\text{guid}})(\hat{y}_i^{\text{pre}} - a_i)\)$ The attenuation coefficient \(\alpha\) is partitioned according to disclosure certainty: retaining 100% displacement (\(\alpha = 1.0\)) under explicit numeric guidance or no-guidance regimes; attenuating displacement by 50% (\(\alpha = 0.5\)) for qualitative-only or forward-looking narrative commentary; and completely nullifying displacement (\(\alpha = 0.0\), collapsing back to \(a_i\)) for weakly derived numeric signals. Concurrently, CAME produces an end-to-end source-linked trace documenting the selected statistical prior, active evidence cards, retrieved temporal analog quarters, and individual expert weight allocations, establishing full auditability for every prediction.
Benchmark & Evaluation Protocol¶
To guarantee empirical validity without look-ahead bias, the evaluation is structured as a 336-row rolling backtest across 12 technology and platform firms (FY2019 Q1–FY2025 Q4). Method design and protocol selection were strictly isolated to AAPL, NVDA, and AVGO through FY2023 Q4; all parameters were frozen prior to running the 96-row temporal held-out evaluation (FY2024 Q1–FY2025 Q4) and the 9-firm company-held-out robustness panel. All comparative baselines operate under identical online forecast-time observation boundaries.
Key Experimental Results¶
Main Results¶
On a development-inclusive rolling backtest of 336 company-quarters spanning 12 major technology and platform enterprises (AAPL, NVDA, AVGO, AMZN, ASML, GOOGL, INTC, META, MSFT, MU, TSLA, ORCL), CAME is evaluated against text retrieval, document-only LLMs, time-series foundation models, and direct prompting baselines. Metrics report cross-company macro means \(\pm\) sample standard deviations.
| Model Category | Method | N | sMAPE ↓ | MAPE ↓ | MAE (B) ↓ | RMSE (B) ↓ | \(R^2\) ↑ | DA ↑ |
|---|---|---|---|---|---|---|---|---|
| Text & Retrieval | Wide-context revenue | 336 | 0.071 ± 0.031 | 0.072 ± 0.034 | 3.04 ± 3.87 | 4.13 ± 5.41 | 0.876 ± 0.136 | 79.3% ± 14.2% |
| Text & Retrieval | CARAG-style retrieval | 336 | 0.083 ± 0.036 | 0.086 ± 0.044 | 2.98 ± 3.52 | 4.03 ± 4.89 | 0.837 ± 0.176 | 81.2% ± 9.3% |
| Document-Only LLM | Transcript-only (GPT-4.1-mini) | 336 | 0.107 ± 0.053 | 0.119 ± 0.072 | 3.64 ± 3.83 | 5.17 ± 5.32 | 0.553 ± 0.636 | 77.8% ± 11.1% |
| Document-Only LLM | Filing-only (10-Q/K) | 336 | 0.212 ± 0.072 | 0.313 ± 0.170 | 9.12 ± 7.12 | 22.49 ± 20.14 | -15.015 ± 29.041 | 64.8% ± 9.0% |
| Time-Series Baselines | TimeGPT-1 | 336 | 0.097 ± 0.047 | 0.097 ± 0.047 | 2.94 ± 3.23 | 3.75 ± 4.05 | 0.842 ± 0.128 | 71.0% ± 15.2% |
| Time-Series Baselines | Chronos-Bolt Base (205M) | 336 | 0.096 ± 0.045 | 0.095 ± 0.046 | 2.95 ± 3.27 | 3.73 ± 4.06 | 0.843 ± 0.114 | 69.1% ± 16.4% |
| Direct Same-Task LLM | History-only | 336 | 0.098 ± 0.057 | 0.094 ± 0.056 | 2.35 ± 2.13 | 3.27 ± 2.99 | 0.826 ± 0.195 | 76.2% ± 10.8% |
| Direct Same-Task LLM | History + Guidance | 336 | 0.066 ± 0.033 | 0.063 ± 0.030 | 2.07 ± 2.07 | 2.75 ± 2.81 | 0.906 ± 0.093 | 81.2% ± 11.4% |
| Direct Same-Task LLM | History + Transcript | 336 | 0.067 ± 0.044 | 0.090 ± 0.120 | 2.55 ± 2.83 | 5.26 ± 6.88 | 0.295 ± 1.600 | 85.5% ± 11.2% |
| Matched Control | Statistical Anchor (SA) | 336 | 0.070 ± 0.036 | 0.067 ± 0.033 | 2.25 ± 2.33 | 2.96 ± 3.10 | 0.903 ± 0.087 | 74.4% ± 20.9% |
| Ours | CAME | 336 | 0.059 ± 0.035 | 0.058 ± 0.035 | 1.81 ± 1.84 | 2.39 ± 2.36 | 0.929 ± 0.067 | 84.0% ± 11.0% |
| Relative Gains | vs Statistical Anchor | 336 | +15.3% | +13.1% | +19.6% | +19.3% | +0.026 | +9.6 pp |
| Relative Gains | vs History + Guidance | 336 | +9.6% | +7.6% | +12.6% | +13.3% | +0.023 | +2.8 pp |
Ablation Study¶
Ablation experiments conducted over the 336-row development-inclusive backtest, the 96-row strictly temporal held-out evaluation window (FY2024 Q1–FY2025 Q4), and specific guidance strata are summarized below:
| Model Variant | Full (336) sMAPE ↓ | Held-out (96) sMAPE ↓ | Explicit (165) sMAPE ↓ | Non-exp. (65) sMAPE ↓ | No-guid. (106) sMAPE ↓ | Full MAE (B) ↓ | Direction Acc. DA ↑ |
|---|---|---|---|---|---|---|---|
| CAME (Full) | 0.0593 | 0.0378 | 0.0369 | 0.0981 | 0.0857 | 1.81B | 84.0% |
| w/o CurrentEvidence + TemporalMemory | 0.0683 | 0.0519 | 0.0381 | 0.0994 | 0.1276 | 2.25B | 79.9% |
| w/o CurrentEvidence | 0.0671 | 0.0463 | 0.0375 | 0.1011 | 0.1267 | 2.16B | 79.9% |
| w/o TemporalMemory | 0.0659 | 0.0458 | 0.0380 | 0.0954 | 0.1246 | 2.10B | 80.6% |
| w/o AnchorMemory | 0.0654 | 0.0405 | 0.0370 | 0.0981 | 0.1379 | 1.98B | 81.8% |
| w/o ExplicitGuidance | 0.0616 | 0.0386 | 0.0408 | 0.0981 | 0.0857 | 2.00B | 82.7% |
Key Findings¶
- Dominant Impact in No-Guidance Regimes: CAME achieves its most dramatic performance gains on the 106 quarters lacking management guidance, driving sMAPE down from 0.1442 (matched anchor) to 0.0857, reducing MAE from 3.453B to 2.683B, and lifting directional accuracy from 60.0% to 77.9% (+17.9 percentage points). Ablating AnchorMemory causes the no-guidance sMAPE to spike severely from 0.0857 to 0.1379, proving the indispensable value of systematic error correction.
- Synergy of Semantic and Temporal Experts: Jointly removing Current Evidence and Temporal Memory increases overall sMAPE from 0.0593 to 0.0683 and reduces DA from 84.0% to 79.9%. In the Non-Explicit Guidance slice, however, removing Temporal Memory slightly improves sMAPE (0.0981 to 0.0954), indicating that qualitative forward-looking comments and retrieved analogs can occasionally introduce redundant signals that over-amplify adjustments.
- Catastrophic Failure of Free Document Prompting: Unconstrained document-only models (Filing-only and MD&A-only) completely disintegrate, posting negative \(R^2\) values (-15.015 and -63.910) and colossal RMSE errors (22.49B and 14.07B), demonstrating that raw textual generation without scale-anchored priors cannot reliably perform enterprise-scale financial calculations.
Highlights & Insights¶
- Residual Refinement as a Grounded Paradigm: By reframing LLM reasoning into an anchor-relative residual adjustment task (\(\hat{\delta}_i\)), CAME delegates baseline revenue scale preservation to econometric anchors while leveraging language models solely for marginal semantic corrections.
- Guidance-Gated Decoupling: Structuring the pipeline to conditionally activate memory corrections only on no-guidance rows prevents double-counting information already priced into management disclosures, ensuring clean statistical interpretation.
- Unified Logic with Externalized Company Schemas: The entire reasoning architecture, thresholds, and guardrails are shared identically across all 12 diverse corporate giants (ranging from semiconductor foundries to e-commerce and automotive platforms), demonstrating strong cross-firm representational decoupling without per-company parameter overfitting.
Limitations & Future Work¶
- Evaluation Panel Concentration: The empirical study evaluates 12 mega-cap technology and platform enterprises with dense analyst coverage and detailed disclosures; evaluating small-cap firms, volatile biotech sectors, and companies undergoing frequent segment reorganizations remains an unaddressed frontier.
- Anchor Inertia During Sharp Regime Shifts: When a statistical anchor initializes significantly above the true revenue target and forecast-time text lacks sufficiently decisive downward counter-evidence, the final prediction can remain biased upwards (as observed in the TSLA FY2025 Q4 boundary failure case).
- Association vs. Causal Provenance: While source-linked evidence cards provide clear provenance regarding which textual claims influenced the prediction, the traces represent associative feature provenance rather than mathematically verified causal mechanisms.
Related Work & Insights¶
- vs Time-Series Foundation Models (TimeGPT-1 / Chronos-Bolt): Pretrained numerical time-series models operate with zero narrative awareness, missing critical forward-looking signals (e.g., product launch timelines or executive guidance) that appear in text before manifesting in quarterly lags; CAME grounds narrative signals upon statistical foundations.
- vs Direct LLM Document Prompting (DocFinQA / CARAG): Conventional prompt-based financial NLP models attempt end-to-end numerical synthesis directly from text tokens, frequently generating mathematically implausible revenue swings; CAME constrains generative reasoning to bounded log-residual spaces with strict revenue guardrails.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates an innovative anchor-relative residual paradigm that successfully tames generative hallucination in corporate fundamental forecasting.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Features an exemplary evaluation design with 336-row rolling backtests, 96-row temporal holdouts, granular guidance stratifications, and paired bootstrap significance bounds.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally rigorous and transparent, providing detailed mathematical formulations, diagnostic failure boundary analyses, and realistic cost profiles.
- Value: ⭐⭐⭐⭐⭐ Sets a new methodological benchmark for grounded, verifiable numerical reasoning in high-stakes financial applications.