Neural Computing and Applications (2023) 35:15109–15128

https://doi.org/10.1007/s00521-023-08528-7 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Agricultural commodity price prediction model: a machine learning

framework
Manas Kumar Mohanty1 • Parag Kumar Guha Thakurta1 • Samarjit Kar2

Received: 17 August 2022 / Accepted: 21 March 2023 / Published online: 6 April 2023
 The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2023

Abstract
An efficient machine learning-based framework for crop price prediction is proposed in this paper to assist the farmers in
estimating their profit-loss beforehand. The proposed work is composed of four functional blocks, such as crop yield
prediction, determination of supply, demand prediction and crop price prediction. The input datasets consist of the various
field values, such as yield, remaining crop at the end of the year, import, demand and price of a crop. Various time series-
based algorithms, such as autoregression, moving average, autoregressive moving average, autoregressive integrated
moving average and exponential smoothing, are used to forecast the crop yield. The supply of the crop is determined as a
sum of three variables, i.e., the predicted crop yield, residue and import values. The demand for the crop is predicted from a
year alone as the demand has more correlation with year over other factors. The crop price from demand, supply and year is
predicted using different approaches, which include the time series method, statistical approaches and machine learning
techniques. Finally, these three techniques for price prediction are compared to determine the best model having minimum
root-mean-square error value. In the proposed work, the decision tree regressor is found to be the best model, for predicting
crop price, over others. The superiority of the proposed work over existing approaches, in terms of various aspects, is
shown by simulation results.

Keywords Crop price prediction  Crop yield  Demand and supply  Machine learning  Agriculture

1 Introduction incomes among the poorest compared to other sectors. In

2016, it was found by an analysis [2] that 65% of poor
Agricultural development is one of the most powerful tools working adults made a living through agriculture. How-
to end extreme poverty, boost prosperity, and feed a pro- ever, more than 10% of the world’s population [3] suffered
jected 9.7 billion people by 2050 [1]. Growth in the agri- from food scarcity in 2020. Under such a scenario,
culture sector is two to four times more effective in raising expanding food production is a compelling process to cope
with a shortfall of food with respect to the growing pop-
ulation. In agriculture, crop yield is a measurement of the
Parag Kumar Guha Thakurta and Samarjit Kar have
contributed equally to this work. amount of a crop produced during a particular time period.
Hence, crop yield prediction [4] is most significant for
& Manas Kumar Mohanty global food production. Various factors affecting crop yield
[email protected] are soil type, soil nutrients, temperature, rainfall, etc.
Parag Kumar Guha Thakurta Farmers are benefited from yield forecasts to get assistance
[email protected] in financial and managerial decisions. Hence, an increase in
Samarjit Kar crop yield also improves farmer’s profit, enhancing their
[email protected] socioeconomic status.
1
Department of Computer Science and Engineering, National Meanwhile, accurate crop price forecasting can be
Institute of Technology Durgapur, A-Zone, Durgapur, helpful for farmers to obtain a reasonable price for their
West Bengal 713209, India yield [5]. At the same time, it is very useful for the farmer
2
Department of Mathematics, National Institute of to make better decisions like when to sell their products or
Technology Durgapur, A-Zone, Durgapur, harvest the crop. Such crop price measures the unit value
West Bengal 713209, India

123
15110 Neural Computing and Applications (2023) 35:15109–15128

received by farmers from the domestic market for a specific LOESS are used to forecast the crop yield. The ARIMA
agricultural commodity produced within a year [6]. The method is found to be the best, in terms of minimum
important determinants influencing the crop price are yield, absolute percentage error (MAPE) value, among others
crop type, rainfall, historical prices etc. Another promising used in predicting the crop yield. After that, the supply of
factor in predicting the crop price is the demand [7] which the crop is determined as a sum of three variables, such as
measures the quantity of the crop that consumers can the predicted crop yield, residue and import values. Next,
purchase during a given period. Furthermore, the amount the demand for the crop is predicted from the year alone as
of a crop available for consumers as a supply [8] is the demand has more correlation with year over other
obtained from yield, residue and import, which can affect factors. Furthermore, the crop price from demand, supply
the variation of crop price. Here, the residue is the quantity and year is predicted using different approaches, which
of the crop which remains unsold at the end of the year and include the time series method (ARIMA), statistical
is available for consumption in the next year, while import regression (SR) and various ML techniques such as deci-
denotes the amount of the crop, not produced within the sion tree regressor (DTR), random forest (RF), K-nearest
country, which is brought from another country [9]. neighbor regressor (KNN) and kernel ridge(KR). Finally,
In the early days, crop price prediction was performed these three techniques for price prediction are compared to
by the farmers using their past experience in terms of determine the best model having a minimum root-mean-
historical data [10]. Gradually, much research is being square error (RMSE) value. In the proposed work, the best
done to predict the crop price with growing interest. Pre- model DTR having the least RMSE is found over other
diction of the crop price of agricultural products is also models. The superiority of the proposed work over existing
needed to anticipate the negative impact of future price approaches is shown in various aspects. Hence, the major
changes ahead of time. In a real scenario, every farmer contributions of our work are summarized as follows:
desires to know about his profit if he cultivates a particular
• Crop yield is forecasted using different time series-
crop in the next season. This can improve farmer’s living
based ML algorithms, and subsequently, the best one
standards. Most existing research works on predicting crop
among alternatives is determined in terms of having
price using time series methods. However, supply and
minimum MAPE.
demand are the foremost factors in deciding the price of a
• The amount of supply is determined from predicted
commodity [11]. Hence, the question remains open to
yield, residue data and import data.
predicting the crop price more accurately considering those
• Forecasting of demand is obtained by the most accurate
factors. Usually, a considerable part of forecasting crop
time series prediction method from historical data.
yield and crop price in the agricultural framework cannot
• Predicting the crop price for the next year is obtained by
be delineated in a fundamental stepwise procedure, espe-
using different methods such as time series, statistical
cially with complex, incomplete, ambiguous and strident
and ML techniques and the best one amidst others
datasets [12]. In such a scenario, numerous studies indicate
having a minimum RMSE value is subsequently
that machine learning (ML) algorithms [13] have com-
determined.
paratively an improved potential over conventional statis-
• The superiority of the proposed work over existing
tics in order to develop an agricultural framework with
approaches is shown by various performance compar-
remarkable forecasting ability. Henceforth, it motivates us
ison through simulation results.
to develop an ML-based efficient crop price prediction
model using significant parameters related to it, which can The rest of this paper is organized as follows: Section 2
assist the farmers’ anticipations of substantial price chan- presents the literature survey for the completeness of the
ges in the future and their consequences. proposed work. Section 3 introduces useful preliminaries
An efficient crop price prediction using ML is proposed with respect to the proposed work. Section 4 describes the
here to assist the farmers in estimating their profit-loss proposed methodology. Section 5 covers results from dif-
beforehand. The input datasets for the proposed work ferent models and forecasting the real price data. Finally,
consist of the various field values of the crop (corn) for 47 Sect. 6 concludes the paper.
number of years. The fields of the input datasets are yield,
remaining crop at the end of the year, import, demand and
price of the crop for each year. Initially, the various time 2 Literature survey
series-based ML algorithms, such as exponential smooth-
ing (ES), autoregression (AR), theta model (Theta), The price of different crops is predicted by analyzing the
autoregressive integrated moving average (ARIMA), previous rainfall data [14, 15]. In order to predict the crop
moving average (MA), autoregressive moving average price, several contemporary ML techniques are used [5]. In
(ARMA) and seasonal-trend decomposition (STL) using another work [16], an integrated feature selection of

123
Neural Computing and Applications (2023) 35:15109–15128 15111

ARIMA with computational intelligence approaches has Another prediction model for crop price is discussed
been applied for crop price prediction. Other than the using time series methods [27]. However, this work has yet
ARIMA, the components of the proposed integrated fore- to consider other input parameters except historical prices
casting models include artificial neural networks (ANNs) in order to develop this price prediction model. In [28], a
and support vector regression (SVR). The paper by [12] comparison based on different algorithms, such as ARIMA,
predicts the crop price for the next rotation. This work is SVR and XGBoost, is introduced concerning price pre-
based on finding suitable data models that help in achieving diction. In this context of price forecasting, the authors of
generality for price prediction. Again, different data mining [29] have used optimal lag selection to improve the pre-
techniques were evaluated on different datasets to solve diction performance in terms of time and reduced error.
crop price prediction issues of seasonal crops [17]. Here, However, the significance of supply and demand factors
the equipped marketing information is analyzed for inte- [30] in forecasting agricultural commodity prices was
gration of domestic markets. Subsequently, the price beyond the scope of their work. Similarly, another paper
transmission from markets to farm gates is perceived, [31] predicts the crop price using statistical regression on
which can lead to a sustainable profit. Another paper [18] time series data of monthly price for 14 years. The price
highlights yield and price forecasting for crop decision forecast of a crop using ARIMA is presented by [32].
planning. For yearly yield prediction, they introduce a However, it is restricted only to the prediction of the
weather-based time-dependent dataset. Here, the price is monthly price, which would not be beneficial for the
predicted as a sum of the future prices on a commodity farmers in their season-based agriculture. The work of [33]
basis. compares various models and finds ARIMA as suitable for
In order to make decisions for agricultural marketing, an crop price prediction. An ANN-based approach is used to
intelligent crop price prediction model is described in [19]. predict the future crop price [34].
It is addressed by different ML models which can predict In a nutshell, several procedures to predict crop price
the prices of crops in advance. An adaptive crop price accurately are discussed in various ways. It is observed that
prediction for agriculture applications is proposed by [20]. most of these techniques have used the time series method
Another time series-based approach [21] is proposed to or weather data to predict the price of the crop. However,
explore the crop price and crop yield prediction of selected the utility of supply and demand to predict the crop price is
crops to identify the relevant information with respect to beyond the scope of their work. As per our best knowledge,
the market prices and crop yields. Some of the uncertain no prior research was considered earlier to predict price
conditions, such as climate changes, fluctuations in the using the supply and demand of the crop, even after these
market and flooding, to the agricultural process have been being the prime factors of pricing a crop. Another scope of
addressed here. The research work by [22] discusses the research is still open for further reduction of training and
prediction of crop yield using regression analysis. In this testing errors in developing a crop price prediction model,
work, several influencing factors are mentioned which are which in turn can be beneficial to assist the farmers in
related to crop yield prediction. In [23], a framework for decision-making on their profit. Hence, the work proposed
crop price forecasting is designed by analyzing the time in this paper addresses an efficient ML-based technique to
series data. Here, the major features related to developing obtain the reduced error in price prediction using the sup-
crop price prediction models are the historical weather data ply and demand of that particular crop. Furthermore, the
that influence crop production and transportation and data best alternative training algorithms are determined from the
quality-related features obtained by performing statistical perspective of error reduction in crop price prediction.
analysis. Using ML approaches, another location-aware Hence, this ML-based crop price prediction framework,
crop price prediction is discussed [24]. In this work, the using crop yield, supply and demand in agriculture, is
future price of vegetables is predicted depending on the comprehensively presented next.
suitable location for the profitable production of agricul-
tural products. Forecasting of both crop yield and price
using ML techniques have presented in [25]. The predic- 3 Preliminaries
tion of crop price is mainly dependent on factors, such as
rainfall, temperature, market prices, land area and past crop For a comprehensive presentation of the proposed work,
yield. The authors of this paper also predict the price and the descriptive statistics of the input data is shown next. In
the gain for the next twelve months over the past twelve addition, various performance measures used to highlight
months. Another research work employs intelligent tech- efficiency of the proposed work are introduced.
niques on historical prices to predict the crop price for
different agricultural products [26].

123
15112 Neural Computing and Applications (2023) 35:15109–15128

3.1 Materials 4.1 Crop yield prediction

The dataset (D2-corn) (https://www.ers.usda.gov/data-pro The Block 1 considers an input D2  corn½}yield}. Dif-
ducts/feed-grains-database/feed-grains-yearbook-tables/) ferent time series-based ML algorithms, as shown in Fig. 2,
[35] is obtained from the department of agriculture, United are used to train the model for crop yield prediction in
States (U.S.). The D2-corn was created on Friday, Block 1 which is represented as follows:
December 10, 2021. It consists of the fields such as yield,
Y^Xi ¼ fX ðD2  corn½}yield}:headði  1ÞÞ
residue and import of the crop, corn, year wise for 47 years. ð1Þ
In Table 1, the statistical summary of D2-corn is shown. þ b1X þ erriX ; i=1,2, ..., n
Another dataset, D3-corn, obtained from the same [35] was In (1), Y^Xi is the amount of crop yield predicted by an
also created on Friday, December 10, 2021. The D3-corn algorithm ‘‘X’’ for ith instance, headði  1Þ returns the first
includes the fields year, demand and price for the same 47 (i-1) number of instances, b1X is an additive term for X and
years. It has similar statistics on the crop corn, like D2- errXi is the error in prediction for ith instance of D2 
corn, for 47 years. The statistical summary of D3-corn is corn½}yield} by the algorithm X. In Block 1, time series-
shown in Table 2. It is significant to mention here that the based ML algorithms, such as ES, AR, Theta, ARIMA,
field, year, is common to both D2-corn and D3-corn, so that MA, ARMA and STL as shown in Fig. 2, are used to train
an efficient integration can be performed for successive the prediction model using Eq. (1). Finally, all of these
operations. In Tables 1 and 2, ’N’ denotes the number of
seven predicted values, such as Y^ES , Y^AR , Y^Theta , Y^ARIMA ,
samples in that corresponding dataset. In addition, the
description of all parameters of D2-corn and D3-corn are Y^MA , Y^ARMA and Y^STL , are compared. Subsequently, the
outlined in Table 3 for a better understanding of the pro- best one (Y^best ) for the crop yield prediction is determined
posed work. by the following:
Y^best ¼ min Y^method :mapeðÞ
3.2 Performance measures method2ðES;AR;Theta;ARIMA;MA;ARMA;STLÞ

ð2Þ
The following measures [36], shown in Table 4, are used to
In (2), the min operator returns the algorithm for which
determine the performance of the work proposed in this
MAPE is minimum and mape() returns the MAPE of the
paper.
corresponding algorithm. Thus, Y^best corresponds to one of
the algorithms, addressed earlier, has the minimum MAPE
in prediction over others.
4 Proposed methodology

The proposed work aims to predict the crop price for the
4.2 Determination of supply
next farming season. The proposed methodology is com-
In Block 2, the value of supply is determined from values
prised of four individual functional blocks as shown in
of other parameters such as the amount of predicted crop
Fig. 1. The respective functions of these are-(i) Block 1:
Crop yield prediction, (ii) Block 2: Determination of sup- yield from Block 1, i.e., Y^best , the import value of that crop
ply, (iii) Block 3: Demand prediction and (iv) Block 4: in the current year, i.e., D2  corn½}import} and residue
Crop price prediction. These functional blocks are num- left-over at the end of last year, i.e., D2  corn½}residue}.
bered in a sequence according to the order of their occur- Henceforth, the residue can be expressed as follows:
rence in the proposed methodology. The entire Rn ¼ Sðn1Þ  C ðn1Þ ð3Þ
methodology is described next.
where Sðn1Þ is the supply amount for the crop in ðn  1Þth
year and C ðn1Þ denotes the total amount of the crop con-
sumed in the ðn  1Þth year. Therefore, the supply of the

Table 1 Statistical summary of

Parameter N Mean SD Sum Minimum Maximum
D2-corn
Year 47 1998.5 14 95,928 1975 2021
Yield 47 10,218.2 3102.6 480,255.4 6206 15,454
Residue 47 1351.35 272.4 63,513.45 821 1879
Import 47 82.35 19.48 3870.45 8.2 205.5

123
Neural Computing and Applications (2023) 35:15109–15128 15113

Table 2 Statistical summary of

Parameter N Mean SD Sum Minimum Maximum
D3-corn
Year 47 1998.5 14 95,928 1975 2021
Demand 47 9908.15 2847.35 465,676 5767 14,830
Price 47 3.00149 1.17907 141.07 1.5 6.89

Table 3 Description of parameters used in D2-corn and D3-corn 4.3 Demand prediction
Parameter Description
It is already discussed that the demand has a positive
Year Year correlation with the parameter ‘‘Year’’. So, the demand can
Yield Yearly yield of the crop be predicted using various time series methods. In Block 3,
Residue Remaining crop at the end of the year four methods, such as exponential smoothing (ES),
Import Yearly import of the crop autoregression (AR), moving average (MA) and autore-
Demand Yearly demand of the crop gressive integrated moving average (ARIMA), are used to
Price Yearly price of the crop obtain a predicted value for the demand of the corre-
sponding crop. Each of these methods is briefly introduced
here.
• ES: This method predicts the demand value by using a
crop is determined as a sum of the values of three weighted sum of past observations with exponentially
parameters by the following: decreasing weight for older observations. It can be
Sn ¼ Y^best þ I n þ Rn ð4Þ expressed by the following:

where Sn is the supply amount for nth year and Y^best is the D^nES ¼a  D3  corn½}demand}:headðn  1Þ
ðn1Þ
ð5Þ
predicted crop yield obtained in Block 1. Here, for nth þ ð1  aÞ  DES
year, the parameter I n denotes the import value. It is to be
noted here that the parameter Sn has a significant impact on where D^nES is the predicted demand for nth year, D3 
the crop price prediction. corn½}demand}:headðn  1Þ is the demand data up to
ðn1Þ
ðn  1Þth year, DES is the actual demand of ðn  1Þth
year and a is the smoothing factor having a value from

Table 4 Performance measures

Name of the measure Definition Formula Symbols
Pn
Mean absolute error The average of difference between predicted values
MAE ¼ i¼1
ðY^i Y i Þ Y i = Actual value Y^i = Predicted
(MAE) and actual values n
value n = Number of instances
Pn
Mean square error The average of the square of difference between ðY^i Y i Þ2
MSE ¼ i¼1
n
(MSE) predicted values and actual values
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn i i 2
Root-mean-square The root of the average of the square of difference ðY^ Y Þ
error (RMSE) between predicted values and actual values RMSE ¼ i¼1
n
Pn
Mean absolute The mean of the percentage of difference between ðY^i Y i Þ
MAPE ¼ i¼1
Yi
percentage error predicted values and actual values
(MAPE)
R-squared (R2 ) The proportion of the variance of the output R2 ¼ 1  SSE
SST
SSE = Sum of squares error SST =
variable that is explained by input variable(s) Sum of squares total
Modified version of R2 which also takes into
2
Adjusted R-squared R2Adjusted ¼ 1  ð1Rn‘1
Þðn1Þ ‘ = Number of independent
(R2Adjusted ) account the number of independent variable(s) variable(s)
Theil’s U (U) The accuracy of predictive model over the last Pn 12
ðY^i Y i Þ2
i¼1
observation as forecast n

U ¼ P Pn 12
n
Y2 Y^2
i¼1 þ i¼1
n n

123
15114 Neural Computing and Applications (2023) 35:15109–15128

Fig. 1 Evaluation framework of

proposed methodology

Fig. 2 Crop yield prediction by

different ML algorithms

0 to 1. If the value of a is larger, then the level of X

p

smoothing is reduced. Thus, the value of a close to 1 D^nAR ¼ Uni  D3  corn½}demand}:locðn  iÞ

i¼1
has less smoothing effect and provides greater weight to
past observations, while the value of a closer to zero has þ b2AR þ 2n
ðn1Þ ð6Þ
a greater smoothing effect and emphasizes on DES .
• AR: It uses a series of observations from previous time where Ui is the coefficient of ith lagged demand, ‘p’ is
steps as an input to predict the value of the next time the order of the autoregression, b2AR is an additive term
step. The order of AR is the number of immediately
and D^nAR is predicted demand value of nth year and 2n
preceding values in the series which are used to predict
is the residual.
the value at the present time. The demand value by AR
• MA: It predicts the demand by adding up all the data
is obtained by the following:
values up to a specific time which is further divided by

123
Neural Computing and Applications (2023) 35:15109–15128 15115

the sum of the number of time periods. It can find a approaches. The predicted crop price by these algorithms
constantly updated average demand by considering are obtained by the following equations:
equal weightage to all the past values. This MA method
• For SR:
uses a way to predict the demand from D3  corn as
follows: P^iREG ¼FREG ðD23  corn; a3reg Þ
ð9Þ
1 X n1 þ b3reg þ eireg ; i ¼ 1; 2; :::; n
D^nMA ¼ D3  corn½}demand}:locðiÞ ð7Þ
t i¼n1t
where P^iREG denotes the predicted price for ith instance
of D23  corn, a3reg is the coefficient matrix. Another
where D^nMA is the value of demand for the nth year and
‘t’ is the window size of moving average. term b3reg acts as an additive term in (9) and eireg
• ARIMA: It is based on both AR and MA processes. It denotes the error in prediction of ith instance of
has an ability to generalize for non-stationary series. D23  corn.
The predicted demand by ARIMA is obtained by the • For ARIMA:
following: P^nARIMA ¼U1 D3  corn½}price}:locðn  1Þ þ :::
D^nARIMA ¼/1 D3  corn½}demand}:locðn  1Þ þ ::: þ Up D3  corn½}price}:locðn  pÞ
þ /p D3  corn½}demand}:locðn  pÞ þ h1 kn1 þ ::: þ H1 Kn1 þ ::: þ Hq Knq þ b3ARIMA þ l2
þ hq knq þ b2ARIMA þ l1 ð10Þ
ð8Þ
In (10), P^nARIMA is the predicted price for nth instance,
where D^nARIMA
is the predicted demand for nth instance D3  corn½}price}:locðiÞ is the ith instance, ‘p’ is the
of D3-corn[‘‘demand’’], ’p’ is the order of the autore- order of the autoregression, ‘q’ is the order of the
gression, ’q’ denotes the order of the moving average, h moving average, K is MA parameter, U is AR param-
is MA parameter, / is AR parameter, D3  eter, Pnp are lagged values, Knq are lagged errors of
corn½}demand}:locðn  pÞ are lagged values, knq are moving average, b3ARIMA is an additive term and l2 is
lagged errors of moving average, b2ARIMA is an additive the constant.
term and l1 denotes a constant. • For ML:
Thus, these four time series methods have different pro- P^iX ¼ fX ðD23  corn; a3ml Þ þ b3ml þ eiml ; i ¼ 1; 2; :::; n
cessing techniques, which can lead to distinct values of
ð11Þ
prediction for each. Then, the outputs obtained from the
four methods are compared, and subsequently, the best one where P^X denotes the price prediction using algorithm
(Dnbest ) is selected, which has minimum error for demand ‘X’, a3ml is the coefficient matrix, b3ml is the additive
prediction over others. term and eiml is the error for the ith instance. It is to be
mentioned here that four different ML algorithms, such
4.4 Crop price prediction as DTR, KNN, RF and KR, are used to train over the
training dataset to develop the crop price prediction
The function of crop price prediction for next year (Yn ) in model.
Block 4 is detailed in Fig. 3. Initially, the dataset D3-corn
and the supply (Sn ) are considered as the inputs of Block 4. After the completion of the training phase, the perfor-
The feature ‘‘price’’ from D3-corn is selected. It is noted mance of those training models are evaluated using test
that D3  corn½}price} is partitioned to keep the first 70% datasets, such as D3  corn½}price}:Test and
samples as D3  corn½}price}:Train and the rest 30% D23  corn:Test. Finally, the best one amidst others having
samples as D3  corn½}price}:Test for training and testing, the least error is selected as the proposed crop price pre-
respectively. Then, D3  corn½}price}:Train is used to diction model, Modelproposed , as shown as follows:
train the model by the ARIMA method as a time series Modelproposed ¼ min P^method :rmseðÞ ð12Þ
method2ðREG;ARIMA;MLÞ
data. Meanwhile, the inputs D3-corn and Sn are integrated
to obtain the dataset, D23  corn which is partitioned into In (12), the min operator returns the algorithm for which
two subsets, such as D23  corn:Train and RMSE is minimum and rmse() returns the RMSE of the
D23  corn:Test, in a 70:30 ratio using random sampling. corresponding algorithm. Thus, for any unseen sample
Thus, the D23  corn:Train is used to train the model by comprising of (Sn , Dnbest , Yn ), the proposed model can
using statistical regression (SR) and different ML predict the crop price for the next year more accurately.

123
15116 Neural Computing and Applications (2023) 35:15109–15128

Fig. 3 Workflow of the model

for crop price prediction

5 Experimental results software in the machine, having an Intel I7 processor and

16 GB RAM.
Various simulation results are shown to highlight the effi-
cacy of the proposed work. In this context, the simulation 5.2 Simulation results
setup and the dataset description are shown next.
Figure 4 shows the probability distributions of all input
5.1 Simulation setup parameters of D2-corn and D3-corn. The probability dis-
tributions of the parameters can span over the entire range
In order to simulate the proposed model, two datasets such of the sample values for that parameter. These are shown in
as D2-corn and D3-corn having values for a specific crop, Fig. 4a–e, which correspond to the data of both D2-corn
i.e., corn, are used. Various useful information regarding and D3-corn. More number of these parameters make the
the parameters of the datasets is shown in Table 5. Firstly, distribution close to normal. In contrast, fewer parameters
we have simulated the function of Block 1 to find the yield appeared to be less normally distributed with slight skew-
predicted for the next season. Next, demand is predicted by ness. From this analysis, it is observed that: (i) most of the
Block 3. At last, the crop price is predicted in Block 4. The parameters are well distributed around the mean and very
proposed work is implemented using Python and Origin few parameters have outliers, (ii) the mean, median, stan-
dard deviation, variance and outliers of input parameters
are understood, and (iii) Both D2-corn and D3-corn are
suitable for further analysis as any statistical tests can
Table 5 Parameter description in D2-corn and D3-corn perform better when the input data follows a normal dis-
Name Unit Data type Dimension Attribute type tribution. The scatter matrix with histogram for all the
parameters of D2-corn and D3-corn are shown in Fig. 5a
Yield Million bushel Float 1 Continuous
and b, respectively. These two figures show the scatter
Residue Million bushel Float 1 Continuous plots of each parameter with respect to all of the other
Import Million bushel Float 1 Continuous parameters to highlight the relationship between those.
Demand Million bushel Float 1 Continuous Here, it is observed that all the parameters are linearly
Price USD Currency 1 Continuous related. The histograms of each parameter are also shown
on the diagonal line of scatter matrix, which represents the

123
Neural Computing and Applications (2023) 35:15109–15128 15117

Fig. 4 Box normal plots of

parameters of D2-corn and D3-
corn

frequency distributions of the parameters. The frequency relationship between each variable by simply staring at the
distributions of more number of parameters are nearly parameters of datasets. Figure 6a can indicate the correla-
normal, except a few parameters have positive skewness. tions between each pair of variables for D2-corn. Here, it is
As these variables are normally distributed in both D2-corn shown that most of the variables are highly correlated with
and D3-corn, Pearson’s correlation is preferred over each other. In Fig. 6b, the correlations between each pair of
Spearman’s correlation to highlight the linear relationships variables for D3-corn are shown. Here, a few of the vari-
between the variables. The correlation analysis is signifi- ables are highly correlated with each other.
cant here as it would be very difficult to understand the

123
15118 Neural Computing and Applications (2023) 35:15109–15128

Fig. 5 Scatter matrix with histogram

Fig. 6 Correlation matrix

Figure 7 shows the values of MAPE for all time series- emphasizes the deviations of predicted yield from actual
based ML algorithms used to predict the crop yield. These yields by what percentage of actual yields, so it focuses on
ML algorithms are trained over the training subset of the magnitude of the error. Thus, the ARIMA(4,1,1) is
D2  corn½}yield}, and subsequently, these are tested over selected as the best in terms of reduced MAPE, for crop
the testing subset of the same. The train-test split for D2  yield prediction.
corn½}yield} is 85% and 15%, respectively. The prediction Figure 8 shows the normal distributions of all crop yield
performance of all these training algorithms is measured algorithms used in the proposed work and that of test data.
using the test dataset. It is observed that the MAPE is It shows the spread of the yield and subsequent weightage
highest for ES and lowest for ARIMA(4,1,1). The MAPE of the various values of the yield. No test dataset can show

123
Neural Computing and Applications (2023) 35:15109–15128 15119

the algorithms, are shown in Tables 10 and 11 (in

Appendix), respectively.
Figure 10 shows the normal distribution curves of
demand and supply data obtained from D2-corn and D3-
corn. It is to be noted that both the normal curves are run on
the same historical data from 1975 to 2021. Due to dif-
ferent variations in both demand and supply data, the
curves differ from each other. The normal curve of demand
data has a higher peak than that of supply data, as the
weightage of the mean demand is more than the weightage
of the mean supply. Both the normal curves start at the
same point; however, they end at different points. The
supply is stretched more, as shown here. It means that the
normal curve of supply data has a higher standard deviation
than that of demand data as the supply values at different
Fig. 7 Performance measures in crop yield prediction using different years are more spread out and demand values of each year
time series-based ML algorithms are more closer to the mean. Although the areas under both
curves are nearly equal, however, the overlapping ratio is
only 60.81% due to the difference in mean values of both.
The difference in mean values is mentioned in Table 6. The
yield value under the peak of the curve is called the mean
yield. It is observed from Fig. 10 that the mean value of
supply is slightly higher than that of demand. So, due to
this difference, there is the residue of the crop at the end of
each year. Here, both plots are not perfectly symmetrical
around the mean. The presence of small positive skewness
in both the normal curves indicates that the right tail is
longer than the left tail for both demand and supply. It
occurs due to a rise in both demand and supply for a few
years. During the training phase of the price prediction
model, the supply is calculated by adding all the three
Fig. 8 Normal distribution of crop yield prediction using different parameters, such as yield, residue and import, of D2-corn.
time series-based ML algorithms and test data This supply acts as one of the three inputs for the training
of Block 4. During the prediction of the crop price for next
perfect symmetry around the mean in normal distributions. year by Block 4, the yield (Y^best ) obtained from Block 1 is
Each of these performances has modest positive skewness, added with residue and import values of next year to obtain
which is obtained due to the nature of the test dataset. All Sn . This Sn acts as one of the three inputs for price
of these algorithms have different standard deviations. It is prediction.
to be noted here that all the normal curves are run on 15% The linear fits of demand and supply are shown in
test data of D2  corn½}yield}. The normal curves show Fig. 11, The fit of demand has higher R2 and R2Adjusted
that the performance of ARMA and Theta have the larger
values in comparison with that of supply. It indicates that
peaks than others. Therefore, it shows a higher frequency
the demand prediction can show the variance of demand
of predictions near the mean yield. Finally, ARIMA(4,1,1)
effectively by using year as the only input variable. Here,
is performing best as it shows a similar curve to actual test
the demand has a higher ’R’-value with the year parameter
data. In order to obtain a detailed analysis of individual
compared to that of supply. The demand has a correlation
algorithms mentioned in Fig. 8, the normal distributions to
value of 0.97 with year, as shown in Table 6. So, it is
predict the corn yield against test data are shown in Fig. 9.
preferred to predict demand using time alone. Both plots
The overlapping ratio of ARIMA(4,1,1) with test data is
are fitted over some of the fields of D2-corn and D3-corn.
65.66% which is the highest among all, while the over-
Due to different trends of supply and demand, the fits of
lapping ratio of ARMA is the lowest. In addition, the
both can vary from each other. Figure 12 shows the per-
detailed parameters of the various time series-based ML
formance of the autocorrelation function (ACF) and partial
algorithms used in the crop yield prediction, along with the
autocorrelation function (PACF) for demand at various
values of the coefficients of the best algorithm among all

123
15120 Neural Computing and Applications (2023) 35:15109–15128

Fig. 9 Detailed graphical

analysis of individual crop yield
prediction algorithms

123
Neural Computing and Applications (2023) 35:15109–15128 15121

predicting demand for the time series methods, such as

MA(3), MA(4), ES, AR(1) and ARIMA(1.0.1). All of these
methods are fitted over the entire D3-corn. Then, the pre-
dictions are compared with the actual prices. It is observed
that the MAE is lowest for ARIMA, while the same is
worst for AR. The MA(3) and MA(4) have less error, hence
it shows that the price trend depends upon the recent his-
tory to some extent. The large error of AR(1) shows the
absence of any seasonal pattern in demand. The
ARIMA(1,0,1), which includes both the AR and MA fac-
tors, has the least error in our proposed model. In addition,
the detail of parameters of the various time series-based
ML algorithms used in the demand prediction, along with
the values of the coefficients of the best algorithm among
Fig. 10 Normal distribution of demand vs supply (1975 to 2021) all are shown in Tables 12 and 13 (in Appendix),
respectively.
Table 6 Summary of supply demand analysis Figure 14 shows the ACF and PACF for price prediction
Parameter Yearly demand Yearly supply
at various lags. The price has very high ACF values for
small lags, which shows the price as closely correlated with
Mean 9908.15 11631.19 recent past values. The values of ACF decrease as the lag
Standard deviation 2847.35 2872.44 increases. Again this ACF value increases after some lags,
Pearson correlation 0.968 0.904 the PACF has a significant value at lag 1 only. There is no
2
R 0.937 0.817 understandable pattern in ACF and PACF values, so the
time series method is ineffective for predicting price.
Fig. 15 shows the trend of change in crop price from
1975 to 2021. Initially, during 1994 and 1995, it is
lags. As the demand has very high ACF values for small observed from Fig. 15b that MA(3) of supply is constant.
lags, it shows the demand as nearly correlated with recent However, that demand is increasing. This leads to an
past values. The value of ACF decreases as the lag increase in price, as seen in Fig. 15a. Again, it is seen that,
increases. It shows no seasonal pattern in demand over during the year 2006, there was a sharp transition in the
time, while the PACF has a significant value at lag 1 only. price raising trend. It occurs due to a sudden decrease in
So, it shows that it is better to apply AR while predicting supply with no variation in demand. During 2013, there
demand using time. Figure 13 shows the mean error in was a solid down-trend which happened due to an increase

Fig. 11 Linear fit of supply and demand (1975–2021)

123
15122 Neural Computing and Applications (2023) 35:15109–15128

Fig. 12 ACF vs. PACF of demand (1975–2021)

Figure 16 shows the predicted prices obtained by the

methods such as SR, ARIMA and ML. It is to be noted here
that DTR is used as an ML algorithm. These methods are
trained over 70% of D23-corn, i.e., D23  corn:Train, and
subsequently tested over the remaining 30% test dataset,
i.e., D23  corn:Test. Due to the different predicting
capabilities of these methods, the prediction results are
notably varied from each other. Overall, it is seen that the
SR algorithm deviates more from the actual price and the
DTR algorithm runs most closely with the actual price,
among others. There is a minimum value, pointed out in
Fig. 16, at the year 1986 in actual data, which is best
Fig. 13 Error in time series prediction of demand predicted by DTR. The ARIMA predicts it closer to the
actual price. The detail of values of the coefficients of the
in supply. This down-trend is evident from MA(3) value of ARIMA is shown in Table 14 (in Appendix). The mini-
supply, sma(3), and samples in Fig. 15b. Lastly, Fig. 15b mum value is worst predicted by SR. There is a maximum
shows that supply is decreasing at a higher rate than value of the price in 2012, which is predicted to be iden-
demand after 2017. This leads to an increase in price, tical to the proposed DTR. While SR could not attain this
which is noticeable in Fig. 15a. Hence, the changes in maximum value due to its polynomial terms. Similarly, the
supply and demand affect the crop price, which indicates ARIMA can provide the predicted price lower than the
the significant importance of the proposed work over value obtained by DTR. Finally, it is observed that the
existing works. DTR method provides the predicted crop price value sim-
Table 7 shows the summary of the price field of D3- ilar to the actual price. Thus, DTR is determined as the best
corn. There are a total of 47 samples from the year 1975 to price prediction model over others.
2021. The mean price is 3.00149 USD per bushel. The Figure 17 shows the results of various performance
median and standard deviation (SD) are also shown in this measures in testing the ML algorithms to predict crop
Table. The different measures of the statistical regression price. The performance of test data varies from each other
for predicting price are shown in Table 8. We have run the for different ML algorithms used in the proposed work.
statistical regression up to degree 6. The value of R is The crop price prediction in Block 4 is trained by DTR, RF,
found as 0.6902, which needs to be more significant to fit KNN and KR. The differences in MAE values of all the
well with the input data. algorithms are noticed in Fig. 17. It is seen that DTR has
the least value of MAE as it captures the pattern in the

123
Neural Computing and Applications (2023) 35:15109–15128 15123

Fig. 14 ACF vs. PACF of Price

Fig. 15 Price trend from 1975 to 2021

Table 7 Statistics of price (USD per bushel) Table 8 Summary output of

Parameter name Value
regression
N Mean Median SD Sum Minimum Maximum
R 0.6902
2 0.4764
47 3.00149 2.54 1.17907 141.07 1.5 6.89 R
Standard error 0.8824
Average error 0.6209
training dataset very well among the four algorithms. The
MAE by RF is the maximum among all, making RF the
worst out of other ML algorithms used here for crop price DTR is the best as it fits best to the training dataset among
prediction. The MSE is highest for RF, whereas the lowest these four algorithms. The errors by RF are found to be
is for DTR. The MSE of all the algorithms is highly varied with higher magnitudes. The RMSE is also the highest for
from each other as it calculates upon squared errors. The RF while the lowest for DTR. The RMSE also focuses on

123
15124 Neural Computing and Applications (2023) 35:15109–15128

Table 9 Result comparison with existing works

Method RMSE MAPE

Ranjani Dhanapal et al. [15] 3.8 –

Pandit Samuel et al. [12] 63.8 –
Wiwik Anggraeni et al. [34] – 16.19%
Proposed 0.115 14.02%

Fig. 16 Fit of price prediction using time series, regression, ML

algorithm (proposed), ARIMA and actual

Fig. 18 MAPE in price prediction using different ML algorithms vs.

test dataset size

existing works indicates that our proposed work predicts

Fig. 17 Performance measures in price prediction using different ML
algorithms
the crop price with the least RMSE.
Figure 18 shows the MAPE for all four ML algorithms
used to predict the crop price. Here, 05% of the dataset is
initially used for testing and 95% for training. This process
the magnitude of the error. However, it is lesser than MSE is repeated in such a way that the test dataset increases by
due to the square root terms after squaring. So the RMSE 5% more than the previous test data set size every time.
values of all ML algorithms used here are also far from The process continues up to 95% size of the test dataset.
each other. So, DTR is the best one in the proposed work as The MAPE results are varied for different ML algorithms
it can provide the highest accuracy during prediction. The at any specific percentage of the test dataset. The process is
errors by RF are found to be with high magnitudes also. performed on DTR, RF, KNN and KR. It is observed that
The RMSE values show similar trends to MSE values. The different algorithms perform better at different percentages
second best training algorithm is found to be KNN and the of the test dataset size. The KNN performs best and DTR
results of KR are very near to KNN. In accordance with performs worst when test dataset is less than 20%. It means
RMSE, the performance of the DTR algorithm used in the for smaller test dataset sizes, KNN fits the input data best.
proposed work over existing works for price prediction is When the test dataset size is about 20%-70%, all algo-
shown in Table 9. The DTR algorithm using rainfall data rithms have similar performance. Here, DTR shows a more
[15] is involved in the prediction of crop price. The method steady and least volatile increase in the test dataset. When
applies six different ML algorithms and finds DTR as best the test dataset increases beyond 70%, the spikes can be
in terms of RMSE. The RMSE in this work is 3.8. Another observed suddenly in the case of all four algorithms. It
work of crop price prediction using DTR in [12] obtains an happens because of a decrease in training dataset sizes,
RMSE of 63.8. Comparing the results obtained from such making it difficult to train on the input datasets for all these
algorithms. DTR performs best for a higher percentage of

123
Neural Computing and Applications (2023) 35:15109–15128 15125

the test datasets, and KNN performs worst. It means that the crop price. Simulation results show that the ML
KNN finds it difficult to fit when the training dataset is less. approach is more suitable for crop price prediction than
Finally, DTR is shown as best for crop price prediction in other time series methods and statistical regression tech-
terms of MAPE. The proposed price prediction using DTR niques. A comprehensive result analysis is contributed after
is found to produce a better result of 14.02% than 16.19% postprocessing of data obtained in simulations. Conse-
[34] in terms of MAPE, as highlighted in Table 9. The quently, the convincing results obtained by data analysis
Theil’s U for predicting the crop price using DTR is found can assist the farmers in deciding crop prices, which can
to be 0.456864. So, according to the U statistics measure improve their economy. Thus, the results observed from
mentioned earlier, it is revealed that the proposed model the precision and efficiency tests illustrate the effectiveness
can be considered a better price prediction method. and versatility of the proposed ML algorithms for crop
price prediction. By building an ML-based price prediction
environment, the proposed method makes it feasible for the
6 Conclusion farmer to bring about the crop price prediction with less
expert dependency and minimum prior knowledge. The
An efficient crop price prediction model is proposed to proposed work can assist the farmers and organizations to
predict the crop price using supply, demand and year data. take beneficial decisions ahead of time which can help
As per our best knowledge, no prior research work pre- them to improve their living standards. Experimenting with
dicting the crop price was done using such supply and data prediction through a wide range of ML predictive
demand of that particular crop, even after these being the algorithms can be observed as a basis for decision-making.
prime factors of pricing a crop. So, from the proposed Exploration of more crop price prediction parameters with
framework, a new research direction has emerged where respect to capita income, the health impact of crop, the
the supply and demand of a particular crop are utilized effect of any disastrous event, the price of neighboring
under a single domain toward obtaining crop price pre- countries, future inflation, and availability of alternative
diction. The prediction model for crop pricing is well crops can be included in the current framework to construct
presented as a significant measure of crop yield. However, a more robust working model in the future. Further
the effect of crop yield in predicting crop price needed to improvement in the computing efficiency of the training
be addressed suitably in existing research works in this process is an intriguing option to be concentrated.
price prediction domain. An exhaustive analysis of the
training datasets can provide ample insights to accurately
forecast the required market price of the crop. Through the Appendix
dataset prediction results, it is evident that the price pre-
diction agent administers the process, suggesting that the See Tables 10, 11, 12, 13 and 14.
proposed method can precisely define the characteristics of

Table 10 Input parameters of

Algorithm Input parameters of algorithm
different algorithms for yield
prediction SP P D Q Trend TS SS S Deseasonalize S degree S jump

ES 12 – – – Add – – Auto – – –
AR – 3 – – – –
Theta 12 – – Auto True
ARIMA (4,1,1) – 4 1 1 – –
MA – – 3
ARMA 3 1
STL – – 3 7 1 1
P = number of lagged forecast errors; D = number of differencing needed; Q = number of autoregressive
terms; SP= number of observations in seasonal period; S = smoothing parameter; SS = length of seasonal
smoother; TS = length of trend smoother; S degree = seasonal degree; S jump = seasonal jump

123
15126 Neural Computing and Applications (2023) 35:15109–15128

Table 11 Coefficients and performance summary of the model using ARIMA (4,1,1) for crop yield
Components Coef Std err z P[ j z j [0.025 0.975]

ARIMA (4,1,1) Coefficients of the model for crop yield

ar.L1  0.5124 0.235  2.184 0.029  0.972  0.053
ar.L2  0.3712 0.312  1.191 0.234  0.982 0.240
ar.L3  0.2561 0.261  0.983 0.326  0.767 0.255
ar.L4  0.1349 0.174  0.774 0.439  0.476 0.207
ma.L1  0.9467 0.229  4.134 0  1.396 0.498
sigma2 4.737e?06 1.2e?06 3.956 0 2.39e?06 7.08e?06
ARIMA (4,1,1) Performance summary of the model for crop yield
Ljung-Box (L1) (Q): 0.41 Jarque-Bera (JB): 3.01 Log Likelihood  346.182 –
Prob(Q): 0.52 Prob(JB): 0.22 AIC 704.364
Heteroscedasticity (H): 0.27 Skew:  0.62 BIC 714.189
Prob(H) (two-sided): 0.02 Kurtosis: 3.61 HQIC 707.859
coef Coefficients; std err Standard error; AIC Akaike information criterion; BIC Bayesian information criterion; HQIC Hannan-Quinn infor-
mation criterion

Table 12 Input parameters of different algorithms for demand prediction

Input parameters of algorithm
SP P D Q Trend TS SS S

MA(3) – – – 3 – – – –
MA(4) 4
ES 12 – Add Auto
AR (1) – 1 – –
ARIMA (1,0,1) 1 0 1

Table 13 Coefficients and

Components Coef Std err z P[j z j [0.025 0.975]
performance summary of the
model using ARIMA(1,0,1) for ARIMA(1,0,1) Performance coefficients of the model for price prediction
demand prediction
const 396.3324 273.511 1.449 0.147  139.739 932.404
ar.L1 0.9971 0.028 35.305 0 0.942 1.052
ma.L1 0.1525 0.19 0.803 0.422  0.22 0.525
sigma2 327.5957 56.788 5.769 0 216.293 438.898
ARIMA (1,0,1) Performance summary of the model for demand prediction
Ljung-box (L1) (Q) 4.9 Jarque-Bera (JB): 41.16 Log likelihood  231.42 –
Prob (Q) 0.03 Prob(JB): 0 AIC 470.839
Heteroscedasticity (H) 4.03 Skew: 1.13 BIC 478.721
Prob (H) (two-sided) 0 Kurtosis: 6.68 HQIC 473.87

123
Neural Computing and Applications (2023) 35:15109–15128 15127

Table 14 Coefficients and

Components Coef Std err z P[j z j [0.025 0.975]
performance summary of the
model using ARIMA (1,0,1) for ARIMA(1,0,1) Coefficients of the model for price prediction
price prediction
Const 3.0495 1.538 1.983 0.047 0.035 6.064
ar.L1 0.9486 0.091 10.466 0 0.771 1.126
ma.L1 0.2624 0.137 1.908 0.056  0.007 0.532
sigma2 0.2534 0.045 5.677 0 0.166 0.341
ARIMA (1,0,1) Performance summary of the model for price prediction
Ljung-box (L1) (Q) 0.01 Jarque-Bera (JB): 9 Log Likelihood  40.233 –
Prob(Q) 0.93 Prob(JB): 0.01 AIC 88.466
Heteroscedasticity (H) 4.56 Skew: 0.85 BIC 96.347
Prob(H) (two-sided) 0 Kurtosis: 4.1 HQIC 91.497

Acknowledgements We thank National Institute of Technology 11. Choi J-S, Helmberger PG (1993) How sensitive are crop yields to
Durgapur, India, for providing access to the computational resources price changes and farm programs? J Agricult Appl Econ
and literatures. 25(1):237–244
12. Samuel P, Sahithi B, Saheli T, Ramanika D, Kumar NA (2020)
Data availability The datasets generated during and/or analyzed Crop price prediction system using machine learning algorithms.
during the current study are available from the corresponding author Quest J J Softw Eng Simul
on reasonable request. 13. Medar R, Rajpurohit VS, Shweta S (2019) Crop yield prediction
using machine learning techniques. In: 2019 IEEE 5th interna-
tional conference for convergence in technology (I2CT), pp 1–5.
Declarations IEEE
14. Akshay Prassanna S, Harshanand B, Srishti B, Chaitanya R,
Conflict of interest All authors of this research paper declare that they Kirubakaran Nithiya Soundari S, Kumar VM, Varshitha Chen-
have no conflict of interest. namsetti VG, Maurya PK (2020) Crop value forecasting using
decision tree regressor and models. Eur J Mol Clin Med, 7(2):
2020
References 15. Dhanapal R, AjanRaj A, Balavinayagapragathish S, Balaji J
(2021) Crop price prediction using supervised machine learning
1. Rehman U, Bahar F, Saad A, Bhat M, Singh L, Mahdi S, Mir MS, algorithms. J Phys Conf Ser 1916:012042
Bashir M, Bhat TA, Palmo T (2022) Trends, challenges and 16. Shao YE, Dai J-T (2018) Integrated feature selection of Arima
prospects of underutilized food crops under temperate ecologies: with computational intelligence approaches for food crop price
a review prediction. Complexity 2018:1–17
2. Rob V (2021) Cattaneo a poverty reduction through the devel- 17. Lavanya K, Raguchander T, Iyengar N (2013) Svm regression
opment of inclusive food value chains. J Integ Agricult and Sonn based approach for seasonal crop price prediction. Int J
20(4):964–978 u-and e-Serv Sci Technol 6(4):155–168
3. von Braun J, Afsana K, Fresco LO, Hassan M (2021) Food 18. Kantanantha N, Serban N, Griffin P (2010) Yield and price
systems: seven priorities to end hunger and protect the planet. forecasting for stochastic crop decision planning. J Agricult Biol
Nature Publishing Group Environ Stat 15(3):362–380
4. Tian X, Engel BA, Qian H, Hua E, Sun S, Wang Y (2021) Will 19. Ghutake I, Verma R, Chaudhari R, Amarsinh V (2021) An
reaching the maximum achievable yield potential meet future intelligent crop price prediction using suitable machine learning
global food demand? J Clean Prod 294:126285 algorithm. ITM Web Conf 40:03040
5. Rakhra M, Soniya P, Tanwar D, Singh P, Bordoloi D, Agarwal P, 20. Bayona-Oré S, Cerna R, Tirado Hinojoza E (2021) Machine
Takkar S, Jairath K, Verma N (2021) Crop price prediction using learning for price prediction for agricultural products
random forest and decision tree regression: a review. Mater 21. Dhanasekaran K, Ramprasath M, Sathiyamoorthi V, Poornima N,
Today Proc. https://doi.org/10.1016/j.matpr.2021.03.261 Jayaraj IA (2021) Meta-learning based adaptive crop price pre-
6. Fafchamps M (1992) Cash crop production, food price volatility, diction for agriculture application. In: 2021 5th international
and rural market integration in the third world. Am J Agricult conference on electronics, communication and aerospace tech-
Econ 74(1):90–99 nology (ICECA), pp 396–402. IEEE
7. Bennett AJ, Bending GD, Chandler D, Hilton S, Mills P (2012) 22. Sellam V, Poovammal E (2016) Prediction of crop yield using
Meeting the demand for crop production: the challenge of yield regression analysis. Indian J Sci Technol 9(38):1–5
decline in crops grown in short rotations. Biol Rev 87(1):52–71 23. Jain A, Marvaniya S, Godbole S, Munigala V (2020) A frame-
8. Fischer R, Byerlee D, Edmeades G (2014) Crop yields and global work for crop price forecasting in emerging economies by ana-
food security. Canberra ACT ACIAR, 8–11 lyzing the quality of time-series data. arXiv preprint arXiv:2009.
9. Smith TR (1935) The wheat surplus. Geograph Rev 04171
25(1):107–116 24. Gamage A, Kasthurirathna D (2019) Agro-genius: crop predic-
10. James HS, Sykuta ME (2006) Farmer trust in producer-and tion using machine learning. Int J Innov Sci Res Technol, 4(10)
investor-owned firms: evidence from Missouri corn and soybean 25. Mulla SA, Quadri S (2020) Crop-yield and price forecasting
producers. Agribusiness Int J 22(1):135–153 using machine learning. Int J Anal Exp Modal Anal
12:1731–1737

123
15128 Neural Computing and Applications (2023) 35:15109–15128

26. Brunda SS, Nimish L, Khan A (2020) Crop price prediction using 34. Anggraeni W, Mahananto F, Rofiq MA, Andri KB, Zaini Z,
random forest and decision tree regression Subriadi AP, et al. (2018) Agricultural strategic commodity price
27. Sabu KM, Kumar TM (2020) Predictive analytics in agriculture: forecasting using artificial neural network. In: 2018 International
forecasting prices of arecanuts in kerala. Procedia Comput Sci seminar on research of information technology and intelligent
171:699–708 systems (ISRITI), pp 347–352. IEEE
28. Yuan CZ, Ling SK (2020) Long short-term memory model based 35. Feed Grains: Yearbook Tables. https://www.ers.usda.gov/data-
agriculture commodity price prediction application. In: Pro- products/feed-grains-database/feed-grains-yearbook-tables/.
ceedings of the 2020 2nd international conference on information Accessed: 2010-09-30
technology and computer communications, pp 43–49 36. Chicco D, Warrens MJ, Jurman G (2021) The coefficient of
29. Yuan CZ, San WW, Leong TW (2020) Determining optimal lag determination r-squared is more informative than SMAPE, MAE,
time selection function with novel machine learning strategies for MAPE, MSE and RMSE in regression analysis evaluation. Peer J
better agricultural commodity prices forecasting in malaysia. In: Comput Sci 7:623
Proceedings of the 2020 2nd international conference on infor-
mation technology and computer communications, pp 37–42 Publisher’s Note Springer Nature remains neutral with regard to
30. Gale D (1955) The law of supply and demand. Math Scandi- jurisdictional claims in published maps and institutional affiliations.
navica, 155–169
31. Mishra MK, Sisodia B, Rai V (2019) Forecasting technique of
Springer Nature or its licensor (e.g. a society or other partner) holds
price of potato of Uttar Pradesh. J Pharm Phytochem 8(3):60–62
exclusive rights to this article under a publishing agreement with the
32. Kumari RV, Venkatesh P, Ramakrishna G, Sreenivas A (2019)
author(s) or other rightsholder(s); author self-archiving of the
Chilli price forecasting using auto regressive integrated moving
accepted manuscript version of this article is solely governed by the
average (Arima). Int Res J Agricul Econ Stat 10(2):290–295
terms of such publishing agreement and applicable law.
33. Ohyver M, Pudjihastuti H (2018) Arima model for forecasting the
price of medium quality rice to anticipate price fluctuations.
Procedia Comput Sci 135:707–711

123