Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
57,394 characters · 28 sections · 62 citation commands
A Benchmark of Classical and Deep Learning Models for Agricultural Commodity Price Forecasting on A Novel Bangladeshi Market Price Dataset
Food price volatility poses a persistent challenge to food security, household welfare, and macroeconomic stability across South Asia. In Bangladesh, a nation of over 177 million people where food constitutes a large share of household expenditure herteux2024forecasting, fao2023food, anticipating near-term retail price movements for key agricultural commodities has direct practical consequences. Farmers benefit from forward-looking price signals when planning planting and sales decisions; policymakers require reliable forecasts to activate market intervention mechanisms before supply shocks cascade to consumers; traders and distributors can reduce post-harvest waste through improved logistics. patil2023forecasting explicitly frames accurate agricultural price forecasting as an enabler of Sustainable Development Goal 2 (Zero Hunger), motivating the development of forecasting infrastructure in food-insecure geographies.
Yet despite these stakes, the quantitative forecasting of Bangladeshi agricultural retail prices remains largely unstudied in the machine learning literature. Two gaps are particularly acute. First, no publicly available daily multi-commodity retail price benchmark exists for Bangladesh; research has been confined to single commodities (typically rice) and classical statistical methods hassan2013forecasting, hasan2020ascertaining, imran2022harnessing. Second, because Bangladeshi commodity prices exhibit discrete step-function dynamics---extended periods of stability punctuated by sudden jumps---it is unclear whether approaches designed for smooth time series transfer to this setting. In particular, the widely-used Prophet framework taylor2018forecasting and large-scale Transformer architectures such as Informer zhou2021informer have not been evaluated under these conditions.
This paper addresses both gaps. Our contributions are:
Our central finding is that commodity price forecastability is fundamentally heterogeneous. No single model dominates across all commodities, and the signal-to-noise structure of a commodity's price series---rather than model complexity---is the primary determinant of forecasting accuracy. The released AgriPriceBD dataset (Mendeley Data: \url{https://data.mendeley.com/datasets/bkmxnrn3hn}) and codebase (\url{https://github.com/TashreefMuhammad/Bangladesh-Agri-Price-Forecast}) are intended as infrastructure for future work, enabling researchers to extend, replicate, and build on these baseline results.
Section (ref) reviews related work. Section (ref) describes the dataset and experimental design. Section (ref) presents results. Section (ref) discusses findings. Section (ref) concludes.
Autoregressive time series models have served as the standard baseline for agricultural price forecasting for decades. box2015time established the ARIMA framework, and SARIMA---its seasonal extension---remains competitive on commodity series with stable periodic structure hyndman2018forecasting. In developing economies, SARIMA has been applied to rice prices in Bangladesh hassan2013forecasting, vegetable prices in India dasari2025price, and pulse markets across South Asia mishra2021state. Its principal limitation is linearity, which fails when prices exhibit nonlinear structural breaks or discrete jumps.
Prophet taylor2018forecasting decomposes time series into smooth trend, seasonality, and holiday components using a piecewise-linear model, and has been widely adopted in applied forecasting due to its interpretability. However, its smoothness assumptions are violated by the discrete step-function price dynamics that characterise developing-economy retail markets---a point this paper demonstrates empirically and discusses in Section (ref).
The study of deep learning models for agricultural price forecasting has been quite prevailing in recent years paul2025deep, aslam2024nbeats. Long short-term memory networks hochreiter1997long provided the first effective deep learning approach to sequential modelling. Bidirectional variants have shown improved performance on agricultural and commodity price series across multiple geographies bahar2024dual, nensi2025implementing, dasari2025price. manogna2025enhancing, singh2025cherry conducted a comprehensive deep learning comparison for agricultural commodity price forecasting in India, finding that ensemble and recurrent approaches outperform classical baselines on volatile series. sari2024various compared optimised machine learning techniques across multiple commodity markets, documenting substantial variation in model performance across commodities---consistent with the heterogeneous forecastability finding reported here. patil2023forecasting emphasised the practical connection between forecasting accuracy and food security outcomes in developing economies.
The Transformer architecture vaswani2017attention replaced recurrence with multi-head self-attention. Informer zhou2021informer extended it to very long horizons via sparse attention, designed for industrial datasets with 10{,}000+ observations. PatchTST nie2022time introduced patch-based tokenisation for time series Transformers, substantially improving efficiency on large benchmarks. As we demonstrate, these large-scale architectures require substantially more training data than small-sample agricultural monitoring contexts typically provide.
Fixed sinusoidal positional encoding vaswani2017attention communicates relative sequence order but carries no information about the absolute temporal position of an observation within a seasonal cycle---a critical limitation for harvest-cycle-driven agricultural prices. kazemi2019time2vec proposed Time2Vec, a learnable temporal embedding that combines a linear trend term with learned sinusoidal functions at discovered frequencies. This allows the model to identify dominant periodicities from data rather than assuming them a priori. muhammad2023transformer previously applied a Transformer-based model to the Bangladesh stock market, establishing the feasibility of attention architectures in the Bangladeshi context; the present work extends this to agricultural retail forecasting with an explicit ablation of temporal encoding.
Though there are available datasets in other countries (e.g.\ India) agmarknet, existing work on Bangladeshi commodity markets is narrow in scope. hassan2013forecasting applied SARIMA to wholesale rice prices. hasan2020ascertaining used machine learning for rice price fluctuation analysis, limited to a single commodity. imran2022harnessing incorporated meteorological covariates into rice price prediction, demonstrating the potential value of exogenous features. islam2024comparative applied ML to Aman rice yields, addressing the production side rather than retail prices. The absence of a publicly available daily multi-commodity retail benchmark has likely constrained research activity in this geography.
Across South Asia more broadly, dasari2025price applied SARIMA-LSTM to vegetable price forecasting in India, and mishra2021state analysed pulse production trends using ARIMA. In other developing-economy settings, nensi2025implementing applied LSTM to high-volatility food commodity prices in Indonesia, and sari2024various optimised ML approaches for commodity price prediction in Turkey. None of these studies addresses the Bangladeshi retail market or provides a reusable daily multi-commodity benchmark.
Table (ref) synthesises the characteristics of closely related studies and identifies the specific gaps addressed by this work.
\paragraph{Data source.} The Bangladesh government market monitoring system publishes daily PDF reports recording minimum and maximum retail prices in Bangladeshi Taka (BDT) per kilogram for agricultural commodities at monitored markets. Five commodities were selected based on nutritional significance and consumption prevalence: garlic, chickpea, green chilli, cucumber, and sweet pumpkin. The extraction covers 22 July 2020 to 4 June 2025, yielding 1,779 daily observations per commodity. AgriPriceBD is deposited on Mendeley Data (\url{https://data.mendeley.com/datasets/bkmxnrn3hn}) and all code is available at \url{https://github.com/TashreefMuhammad/Bangladesh-Agri-Price-Forecast}.
\paragraph{Extraction pipeline.} Because no structured digital API exists, an LLM-assisted extraction pipeline was developed. Figure (ref) illustrates the four-stage process.
\paragraph{Data quality.} Four records in the green chilli series (22, 23, 24, and 29 January 2024) contain zero-valued prices inconsistent with surrounding observations ($\approx$62--70 BDT/kg), attributed to government portal outages. These represent 0.22% of the series and were retained as-is with documentation for transparency.
\paragraph{Cross-commodity structure.} Figure (ref) presents the Pearson correlation matrix across all five commodity mid-price series. Garlic and chickpea exhibit the strongest co-movement ($r=0.61$), consistent with both being imported staples subject to common import policy dynamics. Cucumber shows the weakest correlations with green chilli ($r=0.09$) and sweet pumpkin ($r=0.11$), though a moderate correlation with chickpea ($r=0.44$) suggests some shared supply dynamics. Overall cross-commodity correlations are sufficiently low to support univariate modelling as a meaningful baseline.
\paragraph{Summary statistics and stationarity.} Table (ref) reports summary statistics and Augmented Dickey-Fuller (ADF) stationarity test results. Garlic and chickpea are non-stationary, reflecting multi-year price trends. Green chilli, cucumber, and sweet pumpkin are stationary. This heterogeneity motivates the inclusion of both differencing-based (SARIMA) and level-based (deep learning) models.
Daily prices were forward-filled for any isolated missing dates. Each commodity was processed independently as a univariate time series, a design choice supported by the low cross-commodity correlation observed in Figure (ref). A strict temporal split was applied: 80% training (1,423 days), 10% validation (178 days), 10% test (178 days). No shuffling was applied; temporal ordering was preserved throughout to prevent information leakage from future observations. This protocol is consistent with established time series evaluation practice hyndman2018forecasting.
Standard $k$-fold cross-validation is inappropriate for time series data due to temporal dependencies and look-ahead bias hyndman2018forecasting. The test period (May--June 2025) represents a genuine out-of-sample evaluation on the most recently available data. Walk-forward validation over multiple windows is recommended for future work on extended datasets.
Normalisation used MinMax goodfellow2016deep scaling fit exclusively on the training split. All reported metrics are computed on inverse-transformed (original-scale) predictions. Sliding windows of length 90 days were used to construct model inputs, each producing a 14-day forecast.
Seven forecasting approaches were evaluated, spanning two broad families: classical models---Na\"{i}ve Persistence, SARIMA, and Prophet---which rely on statistical or decomposition-based formulations without neural network components; and deep learning architectures---BiLSTM, Vanilla Transformer, T2V-Transformer, and Informer---which learn representations directly from data. All deep learning models used Adam optimisation kingma2014adam, initial learning rate $1\times10^{-3}$, Huber loss, ReduceLROnPlateau scheduling (factor 0.5, patience 10), early stopping (patience 20), and maximum 150 epochs. Random seed 42 was used throughout.
\paragraph{Na\"{i}ve Persistence} Predicts the next 14 days as equal to the last observed price. Zero parameters; serves as the floor baseline for assessing whether model complexity is justified.
\paragraph{SARIMA} Seasonal ARIMA fit per commodity using the Hyndman-Khandakar automatic order selection algorithm via pmdarima hyndman2018forecasting, with weekly seasonal period $m=7$. Rolling expanding-window evaluation over the test period.
\paragraph{Prophet} Configured with Bangladesh-specific holidays (Ramadan, Eid ul-Fitr, Eid ul-Adha, 2020--2025). Default seasonality settings. Prophet's failure mode is analysed in Section (ref).
\paragraph{BiLSTM} Two-layer bidirectional LSTM hochreiter1997long, hidden dimension 64, dropout 0.1 (garlic, chickpea, cucumber) and 0.3 (green chilli, sweet pumpkin).
\paragraph{Informer (preliminary, excluded from main comparison)} The Informer zhou2021informer was evaluated using its standard configuration (e-layers=2, d-layers=1, $d_\text{model}$=64, 4 heads, $d_\text{ff}$=256, factor=5). Results are reported separately in Section (ref).
\paragraph{Vanilla Transformer} Two Pre-LayerNorm encoder layers; 4 attention heads; $d_\text{model}=64$; $d_\text{ff}=256$; dropout $=0.1$ (garlic, chickpea, cucumber) and 0.3 (green chilli, sweet pumpkin). Input sequences are projected from the univariate price signal to $d_\text{model}$ via a linear layer, then summed with fixed sinusoidal positional encodings vaswani2017attention. The last token passes through a linear head to produce the 14-day forecast. Pre-LayerNorm was adopted for training stability on small datasets.
\paragraph{T2V-Transformer (Time2Vec-Enhanced Transformer)} Architecturally identical to the vanilla Transformer with one modification: fixed sinusoidal positional encodings are replaced by Time2Vec learnable temporal embeddings kazemi2019time2vec. Figure (ref) illustrates the architecture comparison. We emphasise that Time2Vec is an existing method proposed by kazemi2019time2vec; the T2V-Transformer here serves as an ablation target to determine whether learnable temporal encoding improves upon fixed sinusoidal PE in this small-sample agricultural setting, rather than as a methodological contribution in its own right.
Time2Vec maps a scalar time input $\tau$ to a $k$-dimensional embedding:
where $\omega_i$ and $\varphi_i$ are learnable parameters. We use $k=32$, with frequencies initialised on a logarithmic scale ($0.01$--$10.0$) to encourage discovery of both weekly and seasonal cycles. The time index $\tau$ is the global position of each observation in the full five-year series normalised to $[0,1]$, enabling the model to learn inter-year seasonal patterns rather than within-window relative positions. The 32-dimensional output is projected to $d_\text{model}=64$ via a linear layer before summation with value embeddings.
Commodity-specific dropout regularisation was applied based on validation loss dynamics: dropout $=0.3$ for green chilli and sweet pumpkin (both exhibiting rising validation loss under the default 0.1), and dropout $=0.1$ for other commodities. This override was applied consistently to all three deep learning models.
MAE provides an interpretable absolute error in BDT/kg. RMSE penalises large errors more heavily and is sensitive to price spike episodes. MAPE enables cross-commodity comparison on a relative scale.
To assess whether performance differences between the T2V-Transformer and vanilla Transformer are statistically significant rather than attributable to sampling variation, we apply the Diebold-Mariano (DM) test diebold1995comparing with the Harvey-Leybourne-Newbold small-sample correction harvey1997testing. The test uses squared forecast errors over the 1,050 test observations (75 rolling windows of 14 days) with a Newey-West HAC variance estimator at lag $h{-}1=13$. A positive DM statistic indicates the T2V-Transformer has lower loss than the vanilla Transformer.
We note that adjacent sliding windows overlap by $\text{seq\_len}{-}1=89$ observations, meaning the effective sample size is closer to 75 independent windows than 1,050 individual timesteps. The Newey-West correction at lag $h{-}1=13$ partially addresses serial correlation within the forecast horizon but does not fully account for inter-window overlap; consequently, marginal significance results should be interpreted with caution. In cases where the Newey-West variance estimator produces a negative value due to strong loss-differential autocorrelation---observed for green chilli in the T2V vs.\ Transformer comparison---we fall back to the unconditional variance as a conservative alternative, and report the resulting statistic with a dagger ($\dagger$) in Table (ref).
STL decompositions for all five commodities are presented in Figures (ref)--(ref) (Appendix (ref)). The residual-to-seasonal (R/S) ratio---reported in Table (ref)---quantifies the relative dominance of unpredictable noise versus exploitable periodicity in each series. This ratio spans from 0.70 (sweet pumpkin) to 1.32 (chickpea) across all five commodities, as reported in Table (ref). Training-scale constraints ($\approx$1,400 windows) are the binding factor regardless of R/S, with BiLSTM the only DL model achieving statistically significant improvement over na\"{i}ve persistence.
Table (ref) presents the complete performance comparison. Forecast plots for all commodities and all models over the May--June 2025 test period are in Figures (ref)--(ref) (Appendix (ref)).
Table (ref) documents the Informer's failure on all five commodities. Training converged in all cases (early stopping triggered within 22--50 epochs), but the resulting predictions are poorly calibrated. The failure mode is not flat-line collapse but rather erratic oscillation: chickpea and green chilli exhibit prediction variance 50$\times$ and 11$\times$ that of the ground truth respectively, indicating the model amplifies noise rather than tracking signal. Garlic and sweet pumpkin reach more reasonable prediction variance (116% and 77% of ground truth) but with systematic accuracy worse than or comparable to na\"{i}ve persistence. Figure (ref) in the appendix illustrates the erratic prediction pattern for garlic.
The failure is architectural rather than a training failure. The Informer's ProbSparse attention mechanism samples a subset of query positions proportional to $O(\log L)$ and applies max-pooling distilling convolutions that halve the sequence length at each encoder layer. On a sequence of length 90, two distilling layers reduce the effective representation to 22 positions. Combined with a training set of $\approx$1,400 windows, the sparsity and distilling assumptions designed for 10{,}000$+$ observation industrial datasets produce degenerate attention patterns that amplify high-frequency noise rather than learning predictive structure. This finding motivates the use of a full-attention lightweight Transformer in the main comparison.
Table (ref) presents the head-to-head ablation between the vanilla Transformer and T2V-Transformer, isolating the temporal encoding contribution. Ablation bar charts across all three metrics are in Figures (ref)--(ref) (Appendix (ref)).
T2V-Transformer degrades relative to the Vanilla Transformer on all five commodities by MAE. The sole exception across any metric is a modest RMSE improvement on cucumber ($\downarrow$0.4%), which is not statistically significant (DM stat $+0.047$, $p=0.962$). T2V-Transformer collapses catastrophically on green chilli ($+146.1\%$ MAE, $p<0.001$) and degrades substantially on garlic ($+151.6\%$ MAE), chickpea ($+253.3\%$ MAE), and sweet pumpkin ($+51.9\%$ MAE). These outcomes are analysed in Section (ref).
Table (ref) presents the Diebold-Mariano test results for the primary ablation comparison: T2V-Transformer (comparison model) against Vanilla Transformer (reference model). A positive DM statistic indicates the T2V-Transformer achieves lower squared forecast error.
Training curves for all deep learning models across all commodities are presented in Figures (ref)--(ref) (Appendix (ref)). All models converge within 25--70 epochs, with early stopping triggered well before the 150-epoch maximum. For cucumber and garlic, the T2V-Transformer shows closely tracking train and validation loss. For green chilli and sweet pumpkin, the default dropout (0.1) produced diverging train-val loss curves from epoch 5, a clear signature of overfitting to noise. Increasing dropout to 0.3 for all deep learning models on these commodities stabilised training substantially, with the Vanilla Transformer achieving the best deep learning performance on sweet pumpkin (MAE 4.17 BDT/kg). For green chilli, the dominant challenge remains irreducible signal noise rather than overfitting.
The central finding of this study is that commodity price forecastability is structurally heterogeneous, consistent with prior research in other contexts plosone2025svmd, makridakis2022m5. The residual-to-seasonal (R/S) ratio from STL decomposition provides a practical prior for predicting both model selection and overfitting risk. The R/S values for retail mid-price---sweet pumpkin (0.70), green chilli (0.74), garlic (0.93), cucumber (1.23), and chickpea (1.32)---all fall below 1.5, suggesting varying degrees of exploitable seasonal structure. However, learnable temporal encoding confers no statistically significant advantage over fixed sinusoidal encoding on any commodity, and T2V degrades by MAE on all five. This indicates that training-scale constraints ($\approx$1,400 windows) are binding regardless of R/S, with BiLSTM the only model achieving statistically significant improvement over na\"{i}ve persistence on garlic (DM $p=0.039$) and cucumber ($p=0.024$), consistent with its recurrent inductive bias at this data scale.
Despite green chilli's low R/S of 0.74 suggesting detectable annual seasonality, na\"{i}ve persistence dominates and T2V causes catastrophic degradation ($+146.1\%$ MAE, $p<0.001$, DM stat $\approx-1.09\times10^{10}$). This reveals a limitation of R/S as a forecastability proxy: the STL seasonal component captures smooth annual cycles, but green chilli price dynamics are driven by discrete threshold events (monsoon disruptions, border closures, cold storage shortages) that are inherently unpredictable from price history. The R/S ratio from STL can understate forecastability difficulty when price dynamics are threshold-driven rather than cycle-driven.
\paragraph{BiLSTM performance on non-stationary commodities.} BiLSTM achieves the best deep learning performance on garlic (MAE 5.34, RMSE 7.01) and chickpea (MAE 1.91, RMSE 2.62), both non-stationary series. Notably, BiLSTM RMSE on garlic (7.01) is the best result across all models including na\"{i}ve persistence (8.04), and BiLSTM is the only DL model with a statistically significant DM advantage over na\"{i}ve on garlic ($p=0.039$) and cucumber ($p=0.024$), consistent with its recurrent inductive bias. A plausible explanation is that the recurrent inductive bias---processing the input window sequentially with multiplicative gating---may generalise better than self-attention at this data scale ($\approx$1,400 training windows), where the attention mechanism has limited context to learn meaningful query-key structure on non-stationary trending series. This interpretation is consistent with observations by manogna2025enhancing that recurrent and attention-based models have complementary strengths depending on series structure, though ablation over data scale would be required to confirm this.
Green chilli merits explicit treatment. The STL decomposition reveals residual amplitudes substantially exceeding the trend component, against a series mean of 87.3 BDT/kg---a signal-to-noise regime in which all univariate temporal models are expected to fail. Price movements are driven by monsoon-related crop failures, border trade disruptions, cold storage constraints, and localised demand spikes that carry no predictive signal in the price history alone. The na\"{i}ve model's superiority (MAE 3.95 BDT/kg) is a feature of the data generating process rather than a model finding. Improving green chilli forecasting would require exogenous features such as rainfall data, import volumes, or cross-commodity price signals imran2022harnessing, which is identified as a priority for future work.
The T2V-Transformer's catastrophic degradation ($+146.1\%$ MAE over the vanilla Transformer, DM stat $\approx-1.09\times10^{10}$, $p<0.001$) on green chilli is interpretable: learnable temporal parameters overfit to noise in the training period, discovering spurious periodicities that generalise poorly. The vanilla Transformer's fixed encoding, incapable of this overfitting, performs substantially better by comparison.
Prophet's failure across all five commodities---MAPE of 29.3% on garlic, 28.0% on green chilli, and 74.6% on sweet pumpkin---is not a model quality issue but a fundamental incompatibility between its assumptions and the data generating process. All five price series exhibit discrete step-function dynamics: prices remain stable for days or weeks, then jump sharply in response to threshold-triggering supply or policy events. Prophet assumes smooth, continuously differentiable trend and seasonal components. Applied to staircase price data, it attempts to fit smooth splines through sharp discontinuities, generating systematic directional bias throughout the forecast horizon.
We acknowledge that Prophet's changepoint_prior_scale parameter controls trend flexibility and could in principle be tuned to partially accommodate discrete price jumps. However, the fundamental incompatibility---smooth spline fitting applied to staircase dynamics---is architectural rather than a tuning artefact, and we expect no parameter configuration to resolve the systematic directional bias on these series.
This is a practically important finding for forecasting practitioners in similar settings across South Asia and other developing economies where retail prices are partially administered or infrequently updated sari2024various. Standard decomposition-based tools require substantial adaptation or replacement in such contexts.
The Informer's failure provides a clear negative result for practitioners considering large Transformer architectures on small agricultural datasets. Rather than producing flat-line predictions, the model generates erratic, noise-amplifying outputs: chickpea prediction variance reaches 4987% of ground-truth variance, indicating the ProbSparse attention patterns are essentially random on this training set size. The ProbSparse attention mechanism and distilling convolutions were designed for sequences with 10{,}000$+$ observations; applied to 90-step windows from a 1,423-day training set, the sparsity and pooling operations cannot learn coherent attention structure and instead transmit noise. This is not a failure of the Transformer paradigm---the lightweight vanilla Transformer trains stably and competitively---but a dataset-scale mismatch with a specific architectural variant. Researchers considering large-scale Transformer architectures for agricultural price forecasting should verify that their training sets are commensurate with the model's data requirements before drawing negative conclusions about the broader architecture class.
\paragraph{Lookback window.} The 90-day lookback is constrained by the evaluation protocol on a five-year dataset. An annual window (365 days) would better capture full harvest-cycle context but would require either a longer series or a smaller test set. Extending data collection is a priority for future work.
\paragraph{Univariate modelling.} All models operate on univariate price series. Incorporating wholesale prices, weather covariates, import volumes, or cross-commodity signals as exogenous features may substantially improve performance, particularly for green chilli imran2022harnessing.
\paragraph{Hyperparameter optimisation.} Hyperparameter tuning was limited to commodity-specific dropout adjustment due to compute constraints (Google Colab free tier). Systematic Bayesian optimisation may further improve deep learning results.
\paragraph{Single temporal split.} Results are reported for a single held-out test period. Walk-forward validation over multiple test windows would provide more robust performance estimates and is recommended for future work on extended datasets hyndman2018forecasting.
\paragraph{Extraction pipeline accuracy.} The dataset was constructed via LLM-assisted parsing of government PDF reports. While the price-range validation described in Section (ref) and the low documented anomaly rate (0.22%) provide indirect quality assurance, extraction accuracy was not formally quantified against a manually-verified holdout sample. For numerical retail price values---the primary dataset content---vision-model error rates are expected to be low given the structured tabular format of the source documents. Researchers extending this pipeline to additional commodities or time periods should consider spot-checking a random sample of extracted records against original PDFs before use.
This paper introduced AgriPriceBD, a novel five-year daily retail price benchmark dataset for five Bangladeshi agricultural commodities, extracted from government PDF reports via an LLM-assisted digitisation pipeline and released publicly to support reproducible research. Using this dataset, we conducted a systematic comparative evaluation of seven forecasting approaches, with formal statistical significance testing and explicit documentation of two failure modes that have not previously been characterised for developing-economy retail markets.
Four principal findings emerge. First, commodity price forecastability is fundamentally heterogeneous: the STL residual-to-seasonal ratio provides a practical prior for model selection, with na\"{i}ve persistence optimal for random-walk commodities. Learnable temporal representations do not provide statistically significant benefits over fixed encoding at the training scales evaluated here; in fact, four of five commodities show statistically significant Transformer superiority over T2V-Transformer ($p<0.001$, DM test), with only cucumber being non-significant ($p=0.962$). Second, Prophet fails systematically across all five commodities, attributable to the incompatibility between its smooth decomposition assumptions and the discrete step-function price dynamics of developing-economy retail markets. Third, the Informer architecture produces erratic, noise-amplifying predictions (prediction variance reaching 4987% of ground truth on chickpea), confirming that sparse-attention Transformers require training sets substantially larger than small-sample agricultural monitoring contexts can provide. Fourth, Time2Vec learnable temporal embeddings do not provide statistically significant improvements over fixed sinusoidal encoding at this training scale on any commodity. T2V degrades performance significantly on green chilli ($+146.1\%$ MAE, DM stat $\approx-1.09\times10^{10}$, $p<0.001$), and on garlic, chickpea, and sweet pumpkin ($p<0.001$). This negative result for T2V is itself a contribution: practitioners in data-scarce agricultural settings should not assume that learnable temporal encoding improves over simpler fixed alternatives.
Future work should expand data collection to enable annual-window context modelling, incorporate exogenous features such as rainfall and import volumes, apply walk-forward validation on extended series, and extend coverage to additional Bangladeshi commodities. AgriPriceBD is deposited on Mendeley Data (\url{https://data.mendeley.com/datasets/bkmxnrn3hn}) and the complete codebase, including model implementations and the full experimental notebook, is released at \url{https://github.com/TashreefMuhammad/Bangladesh-Agri-Price-Forecast} as infrastructure for these extensions.
AgriPriceBD (A Daily Market Price Dataset of Agricultural Commodities of Bangladesh, July 2020--June 2025) is deposited on Mendeley Data (\url{https://data.mendeley.com/datasets/bkmxnrn3hn}). The complete experimental codebase, including model implementations, the full reproducible notebook, and instructions for replicating all reported results, is openly available at:\\ \url{https://github.com/TashreefMuhammad/Bangladesh-Agri-Price-Forecast}