FARE: Factor-Aware Retrieval with Error-Guided Fusion for Retail Forecasting

Chenghao Gu; Zhuokang Shen; Shenghao Wu; Tingxuan Zhu; Weiyi Luo; Haifeng Tang

cs.LGOct 29, 2026

FARE: Factor-Aware Retrieval with Error-Guided Fusion for Retail Forecasting

Authors: Chenghao Gu, Zhuokang Shen, Shenghao Wu, Tingxuan Zhu, Weiyi Luo, Haifeng Tang

Organizations: East China Normal University, Shanghai, China · China Merchants Bank Credit Card Center, Shanghai, China

Abstract

Retail sales forecasting underpins inventory planning, yet promotions and calendar events reshape short-term demand. Although each factor usually has historical occurrences, the complete factor combination of a target week often has few precedents. The challenge is therefore to retrieve informative partial matches and weight them by their reliability, which varies across series. We propose FARE(Factor-Aware Retrieval with Error-Guided Fusion), which ranks historical weeks separately for each known target-week factor and merges the selected weeks into a factor reference under a fixed budget. Guided by shrunk errors from completed development weeks and the current disagreement among candidates, a first gate mixes this factor reference with a shape reference from similar sales histories, and a second gate controls how strongly the mixture corrects a backbone forecast. Both gates remain fixed during testing. On Favorita, M5, Rossmann, and Rohlik, FARE reduces the weighted absolute percentage error (WAPE) of five backbones by 11.46%, 4.06%, 14.63%, and 10.16% on average. With Chronos-2 frozen, FARE also outperforms conventional and existing enhancement methods, including supervised residual correction and retrieval augmentation. Code is available at https://github.com/HousonGu/FARE .

Figures & tables

Figure 1: Previous methods versus FARE (illustrative example). FARE retrieves partial precedents for each known target-week factor and fuses candidates with gates guided by their historical errors.

Figure 2: Overview of FARE. (a) A backbone forecast and two historical references provide three candidate forecasts. (b) Historical candidate errors and forecast-origin features form the gating inputs. (c) Two gates combine the references and control their contribution to the final forecast.

Dataset	Stage	Origins	Origin range	Forecast-day rows
Favorita	Fit	62	2015-02-23 to 2016-06-27	773,388
Favorita	Dev.	24	2016-07-04 to 2016-12-12	299,376
Favorita	Test	25	2017-02-20 to 2017-08-07	311,850
M5	Fit	165	2011-03-28 to 2014-05-19	23,520,098
M5	Dev.	52	2014-05-26 to 2015-05-18	10,539,837
M5	Test	52	2015-05-25 to 2016-05-16	11,053,322

Table 1: Temporal ranges and forecast-day row counts. Fit denotes supervised-backbone fitting and historical references, and Dev. denotes correction-module development.

Method	Favorita		M5		Rossmann		Rohlik
	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$
Conventional methods
ETS [ 12 ]	17.72	85.4	75.47	0.992	25.07	1477	27.00	37.1
SARIMA [ 3 ]	15.94	76.8	77.85	1.023	23.22	1368	26.39	36.3
Ridge [ 11 ]	18.58	89.6	74.99	0.986	18.52	1091	25.80	35.4
XGBoost [ 4 ]	15.63	75.4	73.30	0.964	15.91	937	20.92	28.7

Table 2: Comparison with conventional and conference methods. FARE uses frozen Chronos-2 as its backbone, and TimeFuse, PIR, RAF, and LightGBM residual correction build on the same frozen forecasts.

Variant	Favorita		M5		Rossmann		Rohlik
	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$
Chronos-2 backbone	15.14	73.0	68.31	0.898	21.20	1249	20.59	28.3
FARE (Ours)	13.28	64.0	64.12	0.843	15.70	925	18.54	25.5
Input and reference-branch ablations
w/o historical errors	14.34	69.1	66.97	0.880	16.86	994	19.65	27.0
w/o factor reference	14.37	69.3	66.98	0.880	18.63	1098	20.11	27.6

Table 3: FARE component ablations with Chronos-2.

Figure 3: Reference allocation and component effects in FARE on the final test rows. (a) WAPE reduction from the backbone for alternative combination rules. Best 50:50 reports the lower test WAPE of the fixed 50:50 B/S and B/F blends. (b) Effects of fixing reference composition (

\alpha

) or correction strength (

\beta

). (c) WAPE increase over full FARE after changing allocation inputs. (d) Component effects on all Favorita rows and the promotion–holiday subgroup.

Backbone	FARE	Favorita		M5		Rossmann		Rohlik
		WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$	WAPE $\downarrow$	MAE $\downarrow$
(arXiv 25’) Chronos-2 [ 1 ]	without	15.14	73.0	68.31	0.898	21.20	1249	20.59	28.3
	with	13.28	64.0	64.12	0.843	15.70	925	18.54	25.5
(NeurIPS 17’) LightGBM [ 13 ]	without	13.70	66.0	65.78	0.865	13.01	767	18.60	25.5
	with	11.98	57.7	63.01	0.828	11.98	706	16.49	22.7
(IJF 21’) TFT [ 14 ]	without	13.91	67.1	64.83	0.852	14.57	859	18.48	25.4

Table 4: Backbone forecasts with and without FARE.

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 873 papers

FreshCast: Timely and Complementary Retrieval for Frozen Time Series Forecasters
Oct 6, 2026Chao He, Jianyu Xu, Xinyi Guo +5Non-Stationary Time Series Forecasting Retrieval-Augmented Time Series Forecasting
Aberdeen Institute of Data Science and Artificial Intelligence, South China Normal University, Foshan, China · School of Artificial Intelligence, South China Normal University, Foshan, China · Faculty of Humanities and Arts, Macau University of Science and Technology, Macau, China
World Models, Trees and Foundation Models for Retail Forecasting: Evidence from Four Public Data Sets
Oct 18, 2026Karthik Iyer, Madan DabbeeruTime Series Forecasting Time Series Forecasting with Covariates
RATL: Learning from Retrieved Residuals for Robust Multivariate Time-Series Forecasting
Sep 3, 2026Yuchen He, Yueyang Cang, Zhiyuan Ning +2Multivariate Time Series Forecasting Residual Learning
Department of Automation, Tsinghua University · State Key Laboratory of Hydroscience and Engineering, Tsinghua University

Oct 6, 2026J/K move · Enter open · S save

FreshCast: Timely and Complementary Retrieval for Frozen Time Series Forecasters

Chao He, Jianyu Xu, Xinyi Guo, Ruiqi Liu+4

Aberdeen Institute of Data Science and Artificial Intelligence, South China Normal University, Foshan, China · School of Artificial Intelligence, South China Normal University, Foshan, China · Faculty of Humanities and Arts, Macau University of Science and Technology, Macau, China

Non-Stationary Time Series Forecasting Retrieval-Augmented Time Series Forecasting

Open paper AskPDF arXiv

Retrieval-augmented forecasting for Web and online services supplements a forecaster with the continuations of historical segments that resemble the current context. Most existing methods build the retrieval memory once from the training segment, leaving observations revealed after deployment unavailable as references, and generally do not calibrate how much retrieved information should influence a frozen forecaster. We identify two key determinants of retrieval utility for a frozen forecaster: the timeliness of the retrieved history and its complementarity, i.e., whether the correction it induces aligns with the forecaster's residual errors; this alignment can shift between validation and deployment when the memory becomes stale. We propose FreshCast, a plug-in retrieval framework for frozen forecasters: a non-parametric streaming memory keeps writing in new observations, relational kernel regression weights candidate readouts into a memory forecast, and its correction weight is calibrated in closed form on the validation segment. Under a simplified generative model, we characterize the optimal combination gain through the second-order relation between forecaster error and memory correction, and show that periodic memory information can become redundant once the forecaster captures it. Across seven datasets, ten forecasters, and three input lengths, FreshCast reduces average MSE for every forecaster and input length, by 14.0% and 5.7% at L = 96 and 720, and achieves lower average MSE than the evaluated retrieval-augmented and online updating baselines in their comparison settings. Ablations show that post-training observations are a primary source of the gain: freezing the memory at the end of training removes most of it. For a frozen forecaster, similarity alone is not enough: useful retrieval must stay timely and provide information that the forecaster still misses.

Abstract

Time Series Forecasting

Explore similar work

FreshCast: Timely and Complementary Retrieval for Frozen Time Series Forecasters

World Models, Trees and Foundation Models for Retail Forecasting: Evidence from Four Public Data Sets

RATL: Learning from Retrieved Residuals for Robust Multivariate Time-Series Forecasting