A hybrid patch--variate attention network with reversible instance normalization, a learned market embedding, and a linearautoregressive residual for multi-horizon panel forecasting of paddy prices

Abstract Daily wholesale prices in Indian agricultural markets (\emph{mandis}) are volatile, weakly autocorrelated from one day to the next, and shaped by supply arrivals, seasonality and weather. Forecasts that target the price that carries the common component of price movement. Second, we forecast the $h$-step \emph{change} rather than the level, which recovers genuine directional skill. Third, we couple the endogenous price history with real meteorological covariates (NASA POWER) and pass both through a hybrid model that combines reversible instance normalization, patch- and variate-level attention, a centralised channel-interaction module, and an explicit linear autoregressive path. Against a field of strong baselines---DLinear, PatchTST, iTransformer, LightGBM, LSTM and Bi-LSTM, all trained identically and evaluated in rupee space---the proposed model attains the lowest error at every horizon. Using the Diebold--Mariano test we find the improvement is \emph{not} significant at one day ahead ($p=0.17$) but becomes significant at seven and fourteen days ($p<0.05$ and $p<0.001$), i.e.\precisely at the decision-relevant horizons. Directional accuracy reaches $\sim$60\% against a 45.8\% majority base rate. A component ablation shows that the returns reformulation and the linear autoregressive path account for most of the gain, and that the attention-based components contribute more at longer horizons. We report the horizon where the improvement is not significant as plainly as those where it is. \vspace{0.5em}\noindent\textbf{Keywords:} agricultural price forecasting; multi-horizon time series; panel data; cross-sectional factor; directional accuracy; Diebold--Mariano test