CAME: Company-Aware Evidence-Memory Experts for Interpretable Quarter-Ahead Revenue Forecasting

Ya-Wen Wu; Meng-Fen Chiang; Kuang-Da Wang; Wen-Chih Peng

cs.CLSep 27, 2026

CAME: Company-Aware Evidence-Memory Experts for Interpretable Quarter-Ahead Revenue Forecasting

Authors: Ya-Wen Wu, Meng-Fen Chiang, Kuang-Da Wang, Wen-Chih Peng

Organizations: National Yang Ming Chiao Tung University, Hsinchu, Taiwan

Abstract

Quarter-ahead revenue forecasting requires company-scale numerical accuracy, strict temporal validity, and company-specific interpretation of narrative disclosures. LLMs can distill textual evidence but can produce scale-misaligned forecasts, whereas history-based anchors are stable but miss forecast-time signals such as product transitions, supply constraints, and management guidance. We introduce CAME (Company-Aware Evidence-Memory Experts), a residual-forecasting framework that refines a no-leakage statistical anchor when current semantic evidence and prior error patterns justify an adjustment. On a development-inclusive rolling backtest of 336 company-quarters from 12 large public technology and platform firms, CAME achieves the lowest aggregate point-estimate error among the reported methods, with statistically supported macro-sMAPE gains over the matched Statistical Anchor, and outperforms History + Guidance on all six aggregate metrics. CAME also links adjustments to source-linked evidence cards and guarded memory traces, supporting forecast inspection, provenance, and failure localization.

Figures & tables

Figure 1: CAME Architecture. CAME integrates company configuration, fiscal quarter, historical revenue, and temporally admissible narrative disclosures into typed, source-grounded Evidence Cards . Starting from a Statistical Anchor , CAME forms a Guidance-Gated Base Anchor . Anchor-relative residual corrections are then composed from three Evidence-Memory Experts : explicit guidance, current evidence, and temporal memory. CAME yields a scale-aligned revenue forecast with a source-linked trace that exposes adjustments and historical analogs for inspection.

Adapted Text and Retrieval Baselines
Method	N	sMAPE $\downarrow$	MAPE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	$R^{2}$ $\uparrow$	DA $\uparrow$
Wide-context revenue	336	0.071 $\pm$ 0.031	0.072 $\pm$ 0.034	3.04B $\pm$ 3.87B	4.13B $\pm$ 5.41B	0.876 $\pm$ 0.136	79.3% $\pm$ 14.2%
CARAG-style retrieval	336	0.083 $\pm$ 0.036	0.086 $\pm$ 0.044	2.98B $\pm$ 3.52B	4.03B $\pm$ 4.89B	0.837 $\pm$ 0.176	81.2% $\pm$ 9.3%
Document-Only LLM Baselines
Transcript-only	336	0.107 $\pm$ 0.053	0.119 $\pm$ 0.072	3.64B $\pm$ 3.83B	5.17B $\pm$ 5.32B	0.553 $\pm$ 0.636	77.8% $\pm$ 11.1%
Filing-only	336	0.212 $\pm$ 0.072	0.313 $\pm$ 0.170	9.12B $\pm$ 7.12B	22.49B $\pm$ 20.14B	-15.015 $\pm$ 29.041	64.8% $\pm$ 9.0%

Table 1: Overall Performance on the Development-Inclusive 336-Row Rolling Backtest. Metric rows are company-level mean

\pm

sample standard deviation;

N

is row coverage. MAE/RMSE use company reporting-currency billions (USD except ASML in EUR), so their cross-company aggregates are descriptive rather than single-currency totals. Lower errors and higher

R^{2}

/DA are better. Improvement rows report relative reference-minus-CAME error, absolute

\Delta R^{2}=R^{2}_{\mathrm{CAME}}-R^{2}_{\mathrm{reference}}

, and DA percentage-point changes; positive values favor CAME.

Panel A. Company-equal metrics
Guidance Slice	N	sMAPE	MAPE	$R^{2}$
Explicit Numeric	165	$.0412/.0369\,(+.0043)$	$.0408/.0369\,(+.0039)$	$.9468/.9655\,(+.0187)$
Non-Explicit	65	$.0994/.0981\,(+.0013)$	$.0983/.0977\,(+.0006)$	$-3.2947/-3.4395\,(-.1449)$
No Guidance	106	$.1442/.0857\,(+.0585)$	$.1310/.0814\,(+.0497)$	$.4180/.7342\,(+.3162)$

Table 2: Guidance Availability on the Development-Inclusive 336-Row Backtest. Cells show Anchor/CAME (delta):

\Delta=\mathrm{Anchor}-\mathrm{CAME}

for errors and

\Delta=\mathrm{CAME}-\mathrm{Anchor}

for

R^{2}

/DA, so positive favors CAME. MAE/RMSE use company reporting-currency billions (USD except ASML in EUR). sMAPE, MAPE, RMSE, and

R^{2}

are company-equal (zero-variance slices omitted for

R^{2}

); MAE and DA are pooled. DA uses the previous same-slice actual and excludes each company’s first slice row.

Variant	Full $\downarrow$	Held-out (96) $\downarrow$	Explicit $\downarrow$	Non-exp. $\downarrow$	No-guid. $\downarrow$	F.MAE $\downarrow$	DA $\uparrow$
CAME	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%

Table 3: Ablation Study. Full and Held-out denote the development-inclusive 336-row and temporal held-out 96-row evaluations. The first five metric columns are company-equal macro sMAPE; F.MAE and DA are 336-row mean company metrics, with F.MAE in company reporting-currency billions. Each removal reruns the reduced path and may recompute dependent proposals and weights, so contrasts are interaction-inclusive end-to-end ablations, not isolated or additive effects. Lower error and higher DA are better.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Evaluation set	Companies	Quarters	Role
Initial development period	AAPL, NVDA, AVGO	FY2019 Q1–FY2023 Q4	initial method development
Initial-firm temporal held-out	AAPL, NVDA, AVGO	FY2024 Q1–FY2025 Q4	temporal held-out evaluation
Company-held-out rolling backtest	nine additional firms	FY2019 Q1–FY2025 Q4	company-held-out robustness evaluation
Company-held-out temporal subset	nine additional firms	FY2024 Q1–FY2025 Q4	joint company/temporal held-out evaluation
Development-inclusive rolling backtest	all 12 firms	FY2019 Q1–FY2025 Q4	development-inclusive aggregate evaluation

Appendix

Table A1: Evaluation sets and their roles.

Surface	Method	$N$	sMAPE $\downarrow$	MAPE $\downarrow$	MAE pool $\downarrow$	RMSE macro $\downarrow$	$R^{2}$ $\uparrow$	DA $\uparrow$
Development-inclusive rolling backtest	TimeGPT-1	336	0.097	0.097	2.94B	3.75B	0.842	71.0%
	Chronos-Bolt Base	336	0.096	0.095	2.95B	3.73B	0.843	69.1%
	Statistical Anchor	336	0.070	0.067	2.25B	2.96B	0.903	74.4%
	History + Guidance	336	0.066	0.063	2.07B	2.75B	0.906	81.2%
	CAME	336	0.059	0.058	1.81B	2.39B	0.929	84.0%
Temporal held-out	TimeGPT-1	96	0.096	0.092	3.57B	4.26B	-0.042	76.2%

Appendix

Table A2: Development-inclusive and temporal held-out results.

Proprietary reference
Model	$N$	sMAPE $\downarrow$	MAPE $\downarrow$	MAE pool $\downarrow$	RMSE pool $\downarrow$	$R^{2}$ $\uparrow$	DA $\uparrow$
GPT-4.1-mini	96	0.052072	0.050990	1.741B	2.379B	0.715	84.5%
Open-weight models
Mistral Small 3 (24B)	96	0.050930	0.050471	1.771B	2.547B	0.529	91.7%
Llama 3.1 8B	96	0.063200	0.064114	1.957B	2.874B	0.132	86.9%
Qwen2.5 14B	96	0.061538	0.059264	2.722B	4.735B	0.562	84.5%

Appendix

Table A3: Supplementary History + Guidance results on the 96-row temporal held-out evaluation.

Candidate family	Examples
Recent-history rules	last-quarter naive; seasonal naive; moving average; drift
Classical time-series	ARIMA; ETS
Direct guidance anchors	Guidance midpoint; guidance-history blend; guidance affine calibration
Revenue-history plus guidance regressors	Linear; Ridge; Elastic Net; Random Forest; XGBoost; LightGBM
Time-series with guidance covariate	SARIMAX with guidance

Appendix

Table A4: Fixed candidate family for the matched Statistical Anchor.

Property	Value
Score metric	sMAPE
Lookback window for scoring	most recent 24 prior realized rows for the same company
Minimum prior history	4 prior realized rows before history-based selection is used

Appendix

Table A5: Statistical Anchor selection policy.

Company	Rows	Guidance mix	Statistical Anchor	History + Guidance	CAME	$\Delta$ vs anchor	$\Delta$ vs LLM
AAPL	28	6 / 14 / 8	0.071 / 5.99B	0.079 / 6.77B	0.057 / 4.97B	+0.014	+0.022
NVDA	28	28 / 0 / 0	0.051 / 0.37B	0.071 / 0.80B	0.051 / 0.35B	+0.001	+0.021
AVGO	28	19 / 0 / 9	0.048 / 0.37B	0.028 / 0.26B	0.033 / 0.27B	+0.016	-0.005
AMZN	28	24 / 1 / 3	0.062 / 6.28B	0.046 / 4.93B	0.043 / 4.13B	+0.019	+0.003
ASML	28	27 / 0 / 1	0.055 / 0.25B	0.058 / 0.28B	0.051 / 0.23B	+0.004	+0.007
GOOGL	28	0 / 10 / 18	0.091 / 5.39B	0.065 / 3.67B	0.082 / 4.81B	+0.009	-0.017

Appendix

Table A6: Per-company sMAPE / MAE heterogeneity on the development-inclusive 336-row rolling backtest. Guidance mix lists CAME’s Explicit / Non-explicit / No-guidance categories.

Figure 2: Comparator trajectories for all 12 firms; shading marks the 96-row temporal held-out window and axes use company-local reporting-currency billions.

Setting	Development-Inclusive (336) sMAPE	Temporal Held-Out (96) sMAPE	No-guidance sMAPE	Anchor–CAME (336)
CAME	0.059253	0.037810	0.085681	+0.010697
$M_{\mathrm{AM}}=4$ or $8$	0.059253	0.037810	0.085681	+0.010697
$\tau_{\mathrm{AM}}=0.01$ or $0.04$	0.059253	0.037810	0.085681	+0.010697
$K_{q}=1$	0.059320	0.038634	0.085910	+0.010630
$K_{q}=5$	0.059557	0.038423	0.085607	+0.010393
$T_{q}=0.20$	0.059180	0.037877	0.085624	+0.010770

Appendix

Table A7: One-factor memory and retrieval sensitivity.

Replicate	Strict Full	Broader Full	Full Difference	Strict Held-Out	Broader Held-Out	Held-Out Difference
1	0.059686	0.059129	-0.000557	0.038409	0.038580	+0.000171
2	0.060097	0.059607	-0.000490	0.038368	0.039171	+0.000803
Mean	0.059892	0.059368	-0.000524	0.038389	0.038876	+0.000487

Appendix

Table A8: Evidence-extraction sensitivity.

Stress	Rate	Full $\Delta$	Held-out (96) $\Delta$	Held-out (96) Anchor–CAME	Guidance attenuation
Polarity flip	5%	+0.000105	+0.000341	+0.009019	77.1%
Polarity flip	10%	+0.000202	+0.000500	+0.008860	71.7%
Forward-card omission	5%	-0.000048	+0.000162	+0.009198	69.5%
Forward-card omission	10%	+0.000062	+0.000401	+0.008959	67.2%
False-positive horizon	5%	-0.000007	+0.000351	+0.009009	65.1%
False-positive horizon	10%	+0.000096	+0.000209	+0.009151	66.1%

Appendix

Table A9: Controlled evidence-card error stress.

Dimension	Primary-evaluator exact agreement	Cohen’s $\kappa$	Adjudicated result / system agreement
Source grounding	88/120 (73.3%)	0.282	102 full; 18 partial (overstatement, omitted qualification, or scope/subject mismatch)
Target-company revenue validity	81/120 (67.5%)	0.269	96/120 valid
Revenue direction	64/120 (53.3%)	0.356	System polarity vs. consensus: 86/110 exact; 10 strict reversals; 14 mixed/one-sided mismatches; 10 non-evaluable (2 no-direction; 8 unclear)
Target-quarter support	74/120 (61.7%)	0.443	55 strict; 41 partial; 24 unsupported
Derived card validity	–	–	49 strict; 84 partial-inclusive

Appendix

Table A10: Outcome-blind internal evaluation of 120 evidence cards.

Table 16

Surface	$N$	Companies	Statistical Anchor sMAPE	CAME sMAPE	$\Delta$ sMAPE	Macro 95% CI
Development-inclusive rolling backtest	336	12	0.069950	0.059253	+0.010697	[+0.007463, +0.014067]
Temporal held-out	96	12	0.047171	0.037810	+0.009360	[+0.004235, +0.016219]
Company-held-out temporal subset	72	9	0.046687	0.036616	+0.010071	[+0.003458, +0.018824]
Explicit Numeric Guidance	165	8	0.041196	0.036894	+0.004302	[-0.001523, +0.009701]
Non-Explicit Guidance	65	7	0.099437	0.098112	+0.001324	[-0.000721, +0.003305]
No Guidance	106	11	0.144171	0.085681	+0.058490	[+0.023849, +0.111277]

Appendix

Table A13: Paired-bootstrap uncertainty against the Statistical Anchor.

Panel of Figure 3 — Figure 3: AAPL FY2024 Q4 traceability case.

Panel of Figure 4 — Figure 4: TSLA FY2025 Q4 boundary trace.

Explore similar work

Long-Horizon Forecasting of Complete Financial Statements with Forma
Aug 11, 2026Travis L. Johnson, Jiannan Jiang, Soumyabrata Chaudhuri +3Long-Term Time Series Forecasting Financial Forecasting
University of Texas at Austin, Austin, TX, USA
From Narrative to Auditable Forecasts: A Structured Scaffold for Agentic Forecasting
Sep 5, 2026Yuanpu Cao, Yongkang Du, Yurui Chang +2Probabilistic Forecasting
The Pennsylvania State University
LLM-based Agents for Forecasting and Prediction: Methods, Training, Evaluation, and Applications
Aug 24, 2026Xiaogang Xu, Jiaqi Tang, Jianmin Chen +12LLM Agents LLM-Based Time Series Forecasting
Zhejiang University, China · The Hong Kong University of Science and Technology, Hong Kong · Northwestern Polytechnical University, China +4

Aug 11, 2026, cs.LGJ/K move · Enter open · S save

Long-Horizon Forecasting of Complete Financial Statements with Forma

Travis L. Johnson, Jiannan Jiang, Soumyabrata Chaudhuri, Yihao Chen+2

University of Texas at Austin, Austin, TX, USA

Long-Term Time Series Forecasting Financial Forecasting Forecasting Benchmarks Financial Time Series Probabilistic Time Series Forecasting

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Long-Horizon Forecasting of Complete Financial Statements with Forma

From Narrative to Auditable Forecasts: A Structured Scaffold for Agentic Forecasting

LLM-based Agents for Forecasting and Prediction: Methods, Training, Evaluation, and Applications