The commodity returns are examined based on a year, month and quarter duration, with the help of a sample of commodity spot price indexes, within a specified period. For instance 1980 to 2018. The prediction depends based on the variables like, bond spreads, increase in the cash flow and increase in the industrial production. For the raw materials and for metals indexes, predictability is strongest. Whereas, for foods and textiles is weakest. When it comes to inflation rate, at the monthly horizon such variables either contains little or no predictive power, but at the quarterly and annual horizons, they seem to have clearer predictive power over the commodity spot prices. Our results show that there are no models which can show the accurate prices of the commodities. This results in finding a new predictive model, which can ease the difficulty (Gargano and Timmermann, 2012).
Another technique is made to demonstrate the costs of the commodities, where it withdraws from the fundamental method in that it influences a clear refinement among the examination of transient i.e., less duration and stationary i.e., long duration administrations. Specifically, this enables us to think of a term, explicit drift for the transient procedure though the stationary procedure is principally drift less because of characteristic high unpredictability of the costs of the commodities, aside from a relatively unimportant mean inversion term, Not suddenly, the data used to manufacture the transient procedure depends on something beyond chronicled costs yet considers extra data about the condition of the market. This inspection is undertaken based on a wide collection of commodities and its costs. The price process’ modeling is related with at least one commodity is of central significance not just in the valuation of an assortment of instruments and the subsidiaries related with these products. Additionally, in the detailing of advancement and concord models, resulted in finding an ideal extraction as well as capacity techniques that will undoubtedly include these costs as parameters. In spite of the fact that our general approach is obviously material to an extensive variety of commodities. Despite the fact that the offer of essential commodities in worldwide yield and exchange has declined over the previous centuries, uncertainty in the costs of the products keep on affecting the worldwide financial actions. For some nations, particularly for the developing nations, the essential commodities remain a critical aspect to gain fare profit, and product value development majorly affects the execution of general macroeconomic. For the strategy of macroeconomic planning and for specifications, the product value conjectures are a key contribution (Wets and Rios, 2015).
Today, in predictive analytics there is no model which can accurately predict the prices of the commodities. This research work projects on developing a new model which predicts these prices accurately. For developing a new model any five predictive models with its algorithm are considered and studied, for predicting the prices of the commodities. Forecasting the prices of the commodity with sensible precision is not clear enough, as it creates confusion by their extensive fluctuation. Indeed, also the long-run drift conduct of commodities’ costs has created banter, as encapsulated by the imperative work, who discovered little proof to help the broadly held Prebisch-Singer see that the costs of the essential commodities were on a declining way completed the long term. In any case, discovering some help for little and variable long-run descending patterns in commodities value information is necessary. In spite of the fact that they additionally locate that such patterns are overwhelmed by the reliably high unpredictability of the product costs. The task includes working on commodity price prediction to know the future prices, when the spot prices have gained less consideration. The spot prices are of discrete intrigue but, as they influence the producer cost and, thus results in inflation of price. In addition, the spot as well as the future prices could be required for influencing the comparable hazard premium varieties.
Thus, there is a demand of multiple-time-ahead predictions. The machine learning theory involves to pattern recognition and statistical inference which helps the model to improvise its performance based on its previous experience.
The purpose of this paper incorporates performing multiple predictions of the prices for the agricultural commodities. The prediction is done prior to 1, 2 or 3 months. These predictions mainly help to take better-informed decisions and also help to manage the risks of price.
The objective of this project is to develop a new predictive model. A predictive Model mind map will be provided. The new predictive model with a perfect algorithm will be provided. Further, all the possible limitations of these five models will be discussed, so that the identified limitations are discarded from the proposing new predictive model, which can help in commodity cost prediction and managing the risks of the prices.
The limitation for this research includes lack of practical applications for the highly complicated models, due to the lack of necessary data and data acquisition expenses. On the other hand, the increased volatility in agricultural commodity prices may increase the difficulty of forecasting accurately making the simple methods less reliable and even the more complex forecast methods may not be robust in this new market environment. To overcome these limitations, machine learning (ML) models can be used as an alternative to complex forecast models.
This research ensures to assess the performance of the agricultural commodity price forecasts, depending on the judgment, historical data on price, and the incorporating prices that are implied by the commodity futures. For instance, in a sample at least 15 commodities’ spot and futures prices looks to be non-stationary and forms a co-integrating relation. The spot prices tends to move towards the future prices in a longer period, and the models for error-correction exploits such feature for producing highly effective and accurate forecasts. The analysis also directs that based on the basis statistical as well as the directional accuracy measures, the futures-based models can generate effective forecasts, especially when compared with the historical-data-based models or judgment, especially at longer horizons (Husain and Bowman, 2004).
Thus, it is hypothesized that whether it is possible to develop a model which can accurately predict the prices of the agricultural commodities. The model is designed in a way that it doesn’t contain any limitations.
This section represents the details of selected Five Predictive Models, ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model.
The five selected predictive models (with algorithms) that estimates the commodities prices in future are listed below:
According to (Khalid, Sultana and Zaidi, 2014), this paper’s aim is not to provide any proof of best methods, instead it aims to present the methods which can be applied in the daily life problems. For instance, for the market participants who are related to the agricultural policy, price recording has become very important to evaluate the impact of market on the international events along with the domestic events. The reason is the huge difference in the commodity prices, and this has led various people in the whole world to find a best method which can evaluate and forecast it accurately. The authors in this paper wish to present the techniques which predicts grain’s market price with the help of ARMA and wavelet transformation. Right from 1983, July to 2013 of July, the monthly data which contains nearly three hundred observations are utilized. Once the model’s precisions are checked, using three set of error tests, the conclusion refers to a best method for forecasting the market price of the grain is wavelet forecasting method. The results of the error test represents that the wavelet is highly important, effective and is feasible based on practical terms. Therefore, the wavelet forecasting model could be utilized as a favorable method to implement for forecasting the market price of the grain, to meet additionally advanced forecasting precision. Thus, it is observed that, Peter Whittle introduced the basic ARMA model and it was well known after George E. P. Box and Gwilym Jenkins used it. It is one form of model which can be expressed even in the lag operator L form. ARMA is a model which is very much popular model and it is adopted by many to predict and forecast, with the help of time series data. Any other form of autonomous variables are not used by the ARMA model, instead only the historical series of a variables are used for prediction.
ARMA stands for Autoregressive Moving Average. Generally, ARMA has the capacity of describing the behavior of the noisy linear dynamical system. Moreover, because of its flexible modeling capability it is capable of representing various unique types of time series. However, regardless of ARMA’s significant success, it assumes that the underlying model is linear and this obstructs its applications with various challenging practical time series. Thus, for resolving this problem, the ARIMA i.e., Autoregressive Integrated Moving Average model is proposed. This is one of the extension of ARMA and it has the capacity to handle nonstationary time series forecasting, with the help of differencing techniques. The purpose of introducing ARIMA method refers to reformulating the ARMA model, for online optimization of complete information, excluding any form of random noise.
The algorithms are explained as follows. A standard game-theoretic framework for online learning is used for the ARIMA model. Here, consecutively the online player commits to a decision and it experiences loss that is not known by the decision maker in his future. This can either be adversarial or it will be based on the actions which the decision maker has taken.
For the ARIMA’s online setting, the coefficient vectors (α, β) are assumed as fixed by the adversary. During the time t, the adversary selects noise t, then it produces resulting observation Xt depending on the Equation 1 and Equation 2. The (α, β)’s true values won’t be revealed to the learner at any time t. Let us assume the iteration of online ARIMA at time t, the learner predicts X˜t, and the true Xt will be revealed to the learner. This results in loss for the learner t(Xt, X˜t).
——Equation 1 and 2
The loss function is defined as follows:
——Equation 3
The online ARIMA learning aims to increase the total losses over certain number of rounds T. The learner’s regret after T rounds is defined as follows:
The aim includes designing an algorithm which is efficient enough to guarantee the regret grows sub linearly as the function T, i.e., RT ≤ o(T). This implies that, the learner’s each round regret can fade away with the increase of T. The Equation 3 mentioned below provides the definition of loss function, which can be considered for applying online convex optimization method for estimating the coefficient vectors (α, β), for the online ARIMA learning task. But, it cannot be achieved as the noise terms {t} are not known to the learner at any time of the online learning process. Thus, despite of (α, β) is provided, it is impossible to perform the prediction, because of the unknown noise terms. Therefore, this challenge is tackled with the help of improper learning principle, for designing the solution where the prediction doesn’t come directly from the original ARIMA model, but from a modified ARIMA model, excluding the explicit noise terms which approximates the original model. Henceforth, the original ARIMA(k, d, q) model with another ARIMA(k + m, d, 0) model (exclusive of noise terms), where m ∈ N is the effectively selected constant like the new ARIMA model with (m + k)-dimensional coefficient vector γ ∈ Rm+k is effective enough for approximating the original prediction:
As a result, the loss function becomes
——Equation 4
The rest of the problem refers to how an appropriate value for parameter m is selected, and what might be the regret with such approximation. As a result a couple of online ARIMA algorithms are focused with the help of two popular online convex optimization solvers like, Online Gradient Descent (ODG) method and Online Newton Step (ONS) (Liu et al., 2016).
Both the algorithms are presented below.
Algorithm 1: ARIMA-ONS(k, d, q)
Input: parameter k, d, m; the learning rate η; the initial (m + k) × (m + k) matrix A0.
Set m = log ((TLMmaxq)−1).
for t=1 to T − 1 do
predict X˜t(γt ) = + ;
receive Xt and incur loss (γt );
Let = ∇ (γt ), update At ← At-1 + ;
Set γt+1← (γt − );
end for
Algorithm 2: ARIMA-OGD(k,d,q)
Input: parameter k, d, q; learning rate η.
Set m = log ((TLMmaxq) −1).
for t=1 to T − 1 do
predict X˜t(γt ) = + ;
receive Xt and incur loss (γt );
Let = ∇ (γt );
Set γt+1← (γt − );
end for
The Algorithm 2 refers to a proposed ARIMA-OGD algorithm, which is used to optimize the coefficient vector with the help of OGD algorithm. When compared to ARIMA-ONS, ARIMA-OGD contains worse regret bound, but computationally it is highly effective.
As per (Kulkarni and Haidar, 2009), the authors talk about the model neural network in this research paper, which depends on the multilayer feedforward network that is used for forecasting the spot price direction of the crude oil, in the short-term like three days in advance. To find the optimal ANN model structure, a lot of concentration was given. However, various method of data pre-processing are tested in this paper. The aim refers to create the benchmark depending on the pre-processed spot price’s lagged value, next the pre-processed futures prices are added for one, two, three and four months to reach maturity, both individually and combined with each other. The optional forecast is found from the results i.e., a dynamic model with 13 lags which for a short-term can forecast the spot price direction. Next, the accuracy of forecast to show the market direction convincingly showed 78 percent, 66 percent and 53 percent for the future 1, 2 and 3 days. The results for every single experiment which contains future data as an input showed that, on the short-term the futures prices surely holds new information on the spot price direction. Hence, the accomplished results can generate comprehensive understanding of the crude oil dynamic which supports the investors and individuals to manage the risks. According to the context of this paper, Artificial Neural Network is the mapping model that is selected. This model is assumed as, nonparametric, nonlinear, assumption free model, which means that it doesn’t perform any prior assumption related to the issue, instead it allows the data to talk by itself. Additionally, it is proved that the feedforward network which contains nonlinear function can approximate any continuous function, then it is used for some time and has achieved success in various studies for different issues along with the price forecasting of the crude oil. In this research paper, the used methodology depends on the three layers feedforward network with the backpropagation algorithm. Here, its goal refers to forecasting the prices of the crude oil for a short-term, and tests whether the future prices of the crude oil contains any new information related to the direction of spot price in the near futures i.e., 3 days in advance. Additionally, whether the information in the futures price is integrated with the spot price can result in providing effectively and comparatively better accuracy of forecast, then it also helps with better overall strategy which creates a benchmark depending on the present and the previous information embedded in the crude oil spot price solely, with the help of three layer feedforward ANN. If this benchmark is created, then the futures prices will be included to measure the performance. For effectively accomplishing this, more focus was given to find the optimal ANN model.
It is observed that, there exists three primary requirements which can make the ANN model successful, they are as follows:
Therefore, the above listed requirements are met successfully, where a wide range of considerations must be considered, followed by data size, data frequency, architect of the network, number of concealed neurons, activation function optimization methods and so on. Even though the guide lines for the design, and certain rules of thump exists. But still, it has no evidence for the existence of such rules that works for the provided issue. Hence, neural network designing is a challenging. The below illustrated flowchart represents a systematic manual method to find the optimal network design for this issue.
Thus, the above flow chart represents that the ANN (Artificial Neural Network) is developed with these main steps such as, selection of performance metrics, data collection, data pre-processing, data normalization, selection of lags, determining the network architecture, selecting the optimization methods, selection of stopping method, in sample training, testing network generalization and out of sample forecasting.
At last, for this research paper the future work will be to investigate the other variable that can help to improve the short term forecast like, the prices of gold, heating oil and interest rate.
While there can be argument with respect to relation among the spot and the futures can vary during the day, i.e., testing with the intraday data can generate varying outcomes. In spite the fact that the crude oil prices’ intra day data isn’t available.
BP Algorithm
The standout amongst the most prominent NN algorithms includes the back propagation (BP) algorithm. It is an algorithm which can be divided into 4 major steps. Subsequent to randomly picking the weights of the network, this algorithm is utilized for registering the vital corrections. The BP algorithm’s four stages are listed below:
Step 1: Feed-forward computation.
Step 2: Output layer’s Back propagation.
Step 3: Hidden layer’s Back propagation
Step 4: Updates of the weight.
When the estimation of the error function has turned out to be adequately little, the algorithm is halted. For BP algorithm, this is the harsh and fundamental formula. There exists certain variety proposed by another researcher, however Rojas definition appear to be very precise and simple for following it. The final step includes updates of the weight occurs all through the algorithm. In the below figure NN comprises of a couple of nodes such as (N0,0 and N0,1) in input layer, a couple of nodes in hidden layer such as (N1,0 and N1,1) and in the output layer a node such as (N2,0) is used. The input layer nodes are associated with the hidden layer nodes which has weights (W0,1-W0,4). The Hidden layer nodes are associated with the output layer node which has weights (W1,0 and W1,1). The provided values to the weights are randomly taken and they are changed when there is iteration of BP. The table with the values of input node and the required output with the learning rate and momentum are likewise presented below the figure. There is additionally sigmoid function formula f(x) = 1.0/(1.0 + exp(−x)). The following indicates the computations for this basic network (Just the estimation for 1st example set is displayed (the input estimations of 1 and 1 with output value 1)). In the training of NN, all the illustration sets are computed yet the logic behind the calculation remains same.
The computation of Feed forward computation comprises of a couple of steps in its process. The initial segment includes taking the estimations of the hidden layer nodes and the following part includes utilizing the hidden layer’s values to register values or output layer’s values. The input values of nodes N0,0 and N0,1 are pushed up to the network to the nodes that are in the hidden layer ( N1,0 and N1,1). They are increased with weights by associating the nodes and calculates the estimations of the hidden layer nodes. The Sigmoid function is utilized for estimations f(x) = 1.0/(1.0 + exp(−x)). N1, 0 = f(x1) = f(w0, 0 ∗ n0, 0 + w0, 1 ∗ n0, 1) = f(0.4 + 0.1) = f(0.5) = 0.622459 N1, 1 = f(x2) = f(w0, 2 ∗ n0, 0 + w0, 3 ∗ n0, 1) = f(−0.1 − 0.1) = f(−0.2) = 0.450166 When the values of the hidden layer are evaluated, the network propagates in the front direction, then from hidden layer it generates the values up to an output layer node (N2,0). This refers to feed forward computation’s second step where, N2, 0 = f(x3) = f(w1, 0 ∗ n1, 0 + w1, 1 ∗ n1, 1) = f(0.06 ∗ 0.622459 + (−0.4) ∗ 0.450166) = f(−0.1427188) = 0.464381
The following step includes computing the error from the node N2,0. By following the below, 1 must be the output. In this example the Predicted value (N2,0) is 0.464381. The computation of error is carried out as follows: N2, 0Error = n2, 0∗(1−n2, 0)∗(N2, 0Desired−N2, 0) = 0.464381(1−0.464381)∗(1−0.464381) = 0.133225. When the mistake is identified, it can be utilized for backward propagation and for the adjustment of weights. This is a two stage process, where the error propagates first from the output layer to the hidden layer. This is the place where for equation, the learning rate and momentum are conveyed to. Thus, the weights W1,0 and W1,1 ae them updated. Prior to updating the weights, the change rate should be identified, which can be finished by multiplying the learning rate, value of error and N1,0 node’s value. ?W1, 0 = β ∗ N2, 0Error ∗ n1, 0 = 0.45 ∗ 0.133225 ∗ 0.622459 = 0.037317 Now, W1,0’s new weight could be evaluated. W1, 0N ew = w1, 0Old + ?W1, 0 + (α ∗ ?(t − 1)) = 0.06 + 0.037317 + 0.9 ∗ 0 = 0.097137 ?W1, 1 = β ∗ N2, 0Error ∗ n1, 1 = 0.45 ∗ 0.133225 ∗ 0.450166 = 0.026988 W1, 1New = w1, 1Old + ?W1, 1 + (α ∗ ?(t − 1)) = −0.4 + 0.026988 = −0.373012. ?(t − 1)’s value is weight’s past delta change. In our case, no past delta change exists thus it is 0 always. On the off chance that, the next iteration are evaluated, this will contain certain value.
Presently, from the hidden layer the errors must be propagated down to the input layer, which is hard when compared to propagating the error from the output layer to the hidden layer. In past case, from the node N2,0 the output known in advance. The output of the nodes N1,0 and N1,1 were not known. First, N1,0 error is found, which is calculated by multiplying the value of new weight W1,0 along with the value of error for the node N2,0. Similar method is used for finding the error for N1,1. N1, 0Error = N2, 0Error ∗ W1, 0N ew = 0.133225 ∗ 0.097317 = 0.012965 N1, 1Error = N2, 0Error ∗ W1, 1N ew = 0.133225 ∗ (−0.373012) = −0.049706. Once the error for the hidden layer nodes is identified, the weights between the input and hidden layer will be updated. For every weight, first the rate of change should be calculated: ?W0, 0 = β ∗ N1, 0Error ∗ N0.0 = 0.45 ∗ 0.012965 = 0.005834 ?W0, 1 = β ∗ N1, 0Error ∗ n0, 1 = 0.45 ∗ 0.012965 ∗ 1 = 0.005834 ?W0, 2 = β ∗ N1, 1Error ∗ n0, 0 = 0.45 ∗ −0.049706 ∗ 1 = −0.022368 ?W0, 3 = β ∗ N1, 1Error ∗ n0, 1 = 0.45 ∗ −0.049706 ∗ 1 = −0.022368. Next, the calculation of new weights takes place among the input layer and the hidden layer. W0, 0N ew = W0, 0Old + ?W0, 0 + (α ∗ ?(t − 1)) = 0.4 + 0.005834 + 0.9 ∗ 0 = 0.405834 W0, 1N ew = w0, 1Old + ?W0, 1 + (α ∗ ?(t − 1)) = 0.1 + 0.005834 + 0 = 0.105384 W0, 2N ew = w0, 2Old + ?W0, 2 + (α ∗ ?(t − 1)) = −0.1 + −0.022368 + 0 = −0.122368 W0, 3N ew = w0, 3Old + ∗?W0, 3 + (α ∗ ?(t − 1)) = −0.1 + −0.022368 + 0 = −0.122368.
It is significant that no weights must be updated unless all the errors are completed calculating. It is anything but difficult to overlook this and if new weights were utilized while computing the errors, the results will not be legitimate. Here is quick second pass with the help of new weights for checking whether there is decline in the error. N1, 0 = f(x1) = f(w0, 0 ∗ n0, 0 + w0, 1 ∗ n0, 1) = f(0.406 + 0.1) = f(0.506) = 0.623868314 N1, 1 = f(x2) = f(w0, 2 ∗ n0, 0 + w0, 3 ∗ n0, 1) = f(−0.122 − 0.122) = f(−0.244) = 0.43930085 N2, 0 = f(x3) = f(w1, 0 ∗ n1, 0 + w1, 1 ∗ n1, 1) = f(0.097 ∗ 0.623868314 + (−0.373) ∗ 0.43930085) = f(−0.103343991) = 0.474186972. After calculating N2,0, the forward pass is finished. The following step includes calculating the N2,0 node’s error. From the below table, 1 must be the output. The predicted value (N2,0) in this case is 0.464381. The error calculation is completed as follows. N2, 0Error = n2, 0∗(1−n2, 0)∗(N2, 0Desired−N2, 0) = 0.474186972∗(1−0.474186972)∗(1−0.474186972) = 0.131102901. Once the initial iteration is completed, the error calculated includes 0.133225 and the newly calculated error denotes 0.131102. This algorithm is thus improved little, however it provides the working of BP algorithm. Although this was very simple example, it should help to understand basic operation of BP algorithm. It can be said that algorithm learned through iterations. The number of iterations in the basic NN might be any number ranging from 10 to 10,000. This is just a single case set pass which can be repeated ordinarily till the point when the errors are little (Cilimkovic, 2011).
According to (Husain and Bowman, 2004), the authors in this research paper examines the three types of commodity price forecasts’ performance. The commodity prices includes, that which depends on the judgment, that which completely relies on the historical price data, and that which incorporates the prices suggested by the commodity futures. In this research work, for most of the fifteen commodities in the sample, spot and futures prices seem to be non-stationary and forms a relation that is co-integrating. Over the long run, the spot prices incline to move towards the futures prices, and the error-correction models which exploits this feature generates highly accurate forecasts. Hence, based on the statistical accuracy measures and directional accuracy measures, this analysis denotes that, the futures-based models can produce more effective forecasts when compared to the historical-data-based models or judgment, mainly at longer horizons. In this study, the overpowering majority of commodity prices are analyzed. It is determined that it has non-stationary characteristics. The commodity prices’ time series properties, spot and futures were checked with the help of unit root tests. The rejection of null hypothesis of a unit root under both the Augmented Dickey Fuller (ADF) test and the Phillips-Perron (PP) test was considered ad the stationarity evidence. Stationarity can’t be rejected just for the soybeans, soybean meal, and soybean oil spot prices. From the futures prices just the prices of the following i.e., tin, maize, wheat and soybean looks as stationary.
It is observed that, the forecasting model which is of easiest form refers to the unit root model which has trend and drift and is presented as follows:
Where, the commodity spot price’s natural logarithm is St, at time t and T is the time trend variable. The error term, et, is considered as the white noise. If the commodity price series comprises of unit root, in such case a difference stationary model must be utilized for modeling the prices. If not, the basic trend stationary model is correct. This easy model could be a helpful benchmark to correlate with the other, more refined models.
The substitute forecasting model can be one which lets an autoregressive process in the first difference of St and for the errors the moving average model. The appropriate time series model for this form includes ARMA model, which can be presented as follows
For obtaining highly accurate forecasts, the futures prices could be included to the unit root model and for the specifications of ARMA. The result of this research work suggests that, the futures prices could deliver realistic guide lines related to developments in spot prices for longer term, at least in directional terms. In this study, for most of the commodities analyzed, the incorporation of futures prices in an error-correction framework generates better and effective forecast performance, for the two-year horizon. The spot and futures prices are co-integrated, and with lower variability in the futures prices, the longer-term spot price movements looks like presented by the futures prices.
It is stated in (Cortazar et al., 2016), that though the commodity pricing models has achieved success to fit the futures prices’ term structure and its dynamics. But, they fail to produce spot prices which are accurate. In this research paper a new method is developed for calibrating these models with the help of not just the futures oil prices observations, but even including the oil spot prices from the analysts´ forecasts. This lets the authors to conclude that, for obtaining the reasonably desired spot curves, the analysts´ forecasts must be utilized either individually or by combining the futures data. The utilization of both the futures and the forecasts, rather than utilizing just the forecasts, creates desired spot curves which extensively don’t vary in the short or the medium period. However, the long term estimations are essentially extraordinary. The incorporation of forecast investigators, along with futures, rather than just the futures costs, doesn’t essentially change the short or medium elements of the futures curve, however it will contain huge impact on the long period estimation of futures. The N-factor Gaussian model is selected for explaining the benefits of utilizing the forecasts of the analysts along with the futures prices. The N – Factor Gaussian model nests various popular commodity pricing models then it provides itself easily to be mentioned without any fixed number of risk factors. For the commodity, the stochastic process of the (log) spot price () is as follows:
It is stated in (Delle Chiaie, Ferrara and Giannone, 2017), that from a broad cross-section of commodity prices, latent factors are extracted in this paper. Based on a single global factor the commodity prices’ bulk fluctuation can be summarized, due to its close relation with global economic activity. The authors state that from the year 2000, the significance of explaining about the variation in the commodity price is very high.
As per (Cortazar et al., 2017), despite the success of the commodity pricing models, in fitting the futures prices’ term structure and its dynamics, it fails to provide accurate and trusted distribution of spot prices. For this reason, the authors have worked on developing a new method for calibrating these models with the help of not just future oil price observations, but also the analysts´ forecasts the oil spot prices. This concludes that, for obtaining the reasonable expected spot curves, the analysts´ forecasts must be utilized, both either alone, or by combining the futures data. Utilizing future data and the forecasts instead of only the forecast can easily provide the expected spot curves which won’t differ in terms of short or medium term. However, when it comes to long term, the estimations show distinct difference.
Thus, the research determines that, in predicting agricultural commodity prices, the model like naïve which is a simple price forecast model, or the distributed-lag models are proved to have achieved well results. Another set of models like, “deferred future plus historical basis” models, autoregressive integrated moving average (ARIMA) models, and then the composite models also provide estimation that is extremely accurate. The fact encountered from the study includes, as the accuracy increases, there will also be increase in the statistical complexity (Ticlavilca, Feuz and McKee, 2010).
The limitation of ARMA and Wavelet model are listed below (Liu et al., 2016):
The limitation of Neural Network Model are listed below (K and Sasithra, 2014):
The limitation of Unit root model are listed below:
The limitation of N – factor Gaussian model are listed below:
Therefore, the above listed limitations from all the selected five models will be discarded for proposing a new model.
In this section a new predictive model along with a brief explanation of the algorithm will be provided, for estimating the prices of the commodities, by considering the discussed five predictive models. Next, the limitations from these five models will be taken out to propose the new model.
The new methodology refers to displaying the prices of the commodities, from the standard method it advances with a refinement among the examination of the transient and stationary (long period) administrations. Specifically, this enables us to think of an unequivocal float term for the transient procedure though the stationary procedure is basically driftless because of innate high volatility of commodity prices, with the exception of a relatively immaterial mean inversion term, Not out of the blue, the data used to assemble the transient procedure depends on something other than chronicled prices yet considers extra data about the condition of the market. This work is done with regards to copper prices however a comparative approach ought to be material to wide assortment of products albeit surely not all since items accompany exceptionally particular qualities. Furthermore, our model likewise considers in?ation which drives us to the multi-dimensional nonlinear system, where unequivocal arrangements can be created (Wets and Rios, 2015).
In this new model, the long period administration will take the qualities of a stationary procedure which will be for the most part in accordance with what can be found in the writing for the ‘in general’ process. Since this is to a substantial degree commonplace domain, we need to get it out of way rather conveniently. The main problem which requires little worry includes choosing whether the model must work with or without any mean inversion and there is extremely no accord which has risen up out of a fairly expound investigation. On one side, essential microeconomics hypothesis says that when prices are high the supply will increment on the grounds that higher cost makers might enter the market which can push the prices down, coming back to the market harmony cost. Then again, if prices are generally low a few makers won’t have the capacity to enter the market and the supply will diminish, invigorating an ascent in prices. The mean inversion hypothesis, presented by, is upheld by numerous creators: demonstrate the presence of mean inversion in spot resource prices of an extensive variety of products utilizing the term structure of future prices; demonstrates a similar utilizing the capacity to fence choice contracts as a proportion of mean inversion; look at commodity prices’ 3 models which considers the mean inversion, and there is numerous different creators that utilization mean returning procedures to display commodity prices.
But, the results demonstrate that the mean inversion is moderate, whereas in others the unit root test neglects dismissing the random walk hypothesis. Further, for instance implement the unit root test on the raw petroleum along with copper prices in the course of recent years. They dismiss the random walk hypothesis, which con?rms that these prices are mean returning. In any case, when they play out the unit root test utilizing the information for just the previous 30 or 40 years, they neglect to dismiss the random walk hypothesis. Clarification which is provided for such an outcome refers to the low speed of inversion, thus utilizing ‘later’ previous information is hard for factually recognizing the mean-returning process along with the random walk. At that point, the reason which one must highly depend on is the hypothetical and conservative consistency (for instance, instinct concerning the activity of equilibrium mechanisms) when compared with the factual tests while choosing which model type is better comparatively. The other illustration is provide, where various models are tested to foresee the medium term of the copper prices ranging between 1 to 5 years, and it infers that the two models which have better execution denotes the following- ?rst-arrange autoregressive process and random walk. The confirmation propose that temporarily (one year) there might be no mean inversion, which is extremely legitimate on the grounds that a maker cannot open all of a sudden another plant, when there are high prices or close to the mine when there are low prices. This is again a contention that backings the method which detaches the lower period and longer period impacts, this depends on the expansive degree on various information base for manufacturing a couple of principle parts of the new model. For long period a stochastic di?erential condition is set up which is the mean returning and, which thusly, will decide the float of the stationary process. Further, it depends on the geometric brownian motion’s variation with the mean inversion that is tuned in with our decision, where it states in?ation free ‘cash’, cf. §5. This model was proposed by [6], and it’s likewise utilized by [14] to display oil prices§. Along these lines, for the stationary process the accompanying system of stochastic di?erential conditions give us the reason for the modeling process:
Where, the is the current estimation of index i, µi and bij are constants that should be evaluated, xt =(,…, refers the system condition at time t, wj,j = 1,…,J are autonomous wiener processes, υi is and file to which returns in the log term and µi is the ‘speed’ at which returns to υi; our procedure will be that this mean-inversion float is moderate and therefore, its in?uence is very lessened.
The following is this system’s solution: for i = 1,…,n,
The solution is being replaced by the term µiυi ds by its expectation. This is proceeded with the introduced error, which is minor and the final calculation of coe?cients µiυi and bij will be very difficult or impossible. Thus, the following stochastic di?erential equation is accepted as a solution for the system: for i = 1,…,n,
By assuming t0 =0, the following is achieved, which is also a log – Gaussian process.
This process’s 1-dimensional version reads as follows,
At last, the properties of Gaussian processes are relied on for calculating the mean as well as the covariance terms of the n-dimensional process. One obtains: for i = 1,…,n,
The mind map represents the selected five predictive model i.e., Bayesian Networks, Decision Trees, Support Vector Machines, Neural Networks and K-Mean.
In /reinforcement learning, Markov decision process is undertaken. Whereas in supervised learning, classification, regression and clustering takes place. Then, to undertake semi-supervised learning summarization takes place. Further for unsupervised learning, anomaly detection is considered as best option. To help with classification and regression, Bayesian networks, Decision trees and support vector machines is utilized. For clustering, neural networks and K-Means is utilized.
But, according to the study, the Predictive models such as, ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model are selected for the same purpose.
The present research has shed light on creating a new predictive model which can accurately predict the prices of the agricultural commodities. The lack of multiple-time-ahead predictions of the prices for the agricultural commodities was the reason of conducting this research. The machine learning theory involves to pattern recognition and statistical inference which helps the model to improvise its performance based on its previous experience. Each objectives of this project are met. That is, a selective five predictive models are selected like, ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model. All these models are considered with its algorithm, for predicting the prices of the commodities. All these models are studied briefly.
It is observed that, the presenting of a value procedure related with at least one commodity is of central significance not just in the estimation of an assortment of instruments and subsidiaries related to these products. Yet additionally, in detailing the advancement and harmony models, resulted in finding an ideal extraction as well as capacity techniques that will undoubtedly include these costs as parameters. In spite of the fact that our general approach is obviously material to an extensive variety of commodities. Despite the fact that the offer of essential commodities in worldwide yield and exchange has declined over the previous centuries, uncertainty in the costs of the products keep on affecting the worldwide financial actions. For some nations, particularly for the developing nations, the essential commodities remain a critical aspect to gain fare profit, and product value development majorly affects the execution of general macroeconomic. Consequently, product value conjectures are a key contribution for the strategy of macroeconomic arrangement and for detailing.
In general, for performing classification and regression the methods like, Bayesian networks, Decision trees and support vector machines are utilized. Then, to support clustering the methods like neural networks and K-Means are utilized. Further, it is determined that, for predicting the prices of the agricultural commodities, the naïve model or the distributed-lag models have accomplished good estimation results. Then, next the models like, “deferred future plus historical basis” models, autoregressive integrated moving average (ARIMA) models, and then the composite models provide highly accurate estimation. The limitation encountered here is, as there is increase in accuracy, the statistical complexity also increases (Ticlavilca, Feuz and McKee, 2010).
The limitations from these five models are taken out to propose a new model. It is recommended that, the large-scale datasets must be managed. Thus, there must not be any failure to manage the large scale datasets. From the other listed limit, it is observed that, the existing algorithms to estimate the ARIMA’s parameter, like least squares and maximum possibility based methods, in advance needs the access of the complete dataset. Hence, this could violate the characteristics of streaming the time series data and fails to manage the problem of concept-drift, which is suggested to improve. Further, there must not be any argument in relation with the spot and futures. Next, the results of unit root must not be slow or fail. Plus, the classification process must be easy and not complicated for a large number of datasets.
Conclusion
This research work has successfully helped to create a new predictive model which can predict the prices of the agricultural commodities accurately. Even an algorithm is provided for this model. Then, all the objectives of this project are met. For instance, a Predictive Model mind map is created with a perfect algorithm. To create this model, five predictive models are selected, such as, ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model. These models are considered with its algorithm, for predicting the prices of the commodities. All these models are studied briefly. Then, all the possible limitations of the selected five models are discussed, which are removed from the proposed new predictive model. The demand of multiple-time-ahead predictions is high which has led to this research, for performing multiple predictions of the prices for the agricultural commodities, prior to 1, 2 or 3 months. The predictions is expected to be helpful in effective decision making and in managing the risks of price. The study determines that, presenting a value procedure linked with even a single commodity has a huge significance and not only in valuation of an assortment of instruments and product’s subsidiaries. Thus, to detail with the advancement and improvement of the models, resulted in finding an ideal extraction as well as capacity techniques, which will undoubtedly include these costs as parameters, in spite of the fact that the general approach is obviously material to an extensive variety of commodities. Despite of the fact that over the previous centuries, the offer of essential commodities in worldwide yield and exchange has declined, the uncertainty in product’s cost has impact on the worldwide financial actions. The main example considered is the developing countries, where the essential commodities remain a critical aspect for gaining fare profit, and the development of the product value highly impacts on the execution of general macroeconomics.
It is believed that the machine learning theory contains pattern recognition as well as statistical inference which ensures to improvise the performance of the model when compared to its previous experience. To perform classification, clustering and regression the methods like, ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model. The study determines a few models like naïve and distributed-lag models have shown effective estimation results to predict the prices of the agricultural commodities. Next, another models include, “deferred future plus historical basis” models, auto-regressive integrated moving average (ARIMA) models as well as the composite models provide very much accurate predictions. Keeping this in mind, one must understand that as the accuracy increases, the statistical complexity also increases.
The algorithms of ARMA and Wavelet model, Neural Network Model, Unit root model and N – factor Gaussian model are studied and discussed, from which the possible limitations are observed and listed.
References
Cilimkovic, M. (2011). Neural Networks and Back Propagation Algorithm. Institute of Technology Blanchardstown.
Cortazar, G., Millard, C., Ortega, H. and Schwartz, E. (2017). Commodity Price Forecasts, Futures Prices and Pricing Models.
Delle Chiaie, S., Ferrara, L. and Giannone, D. (2017). Common factors of commodity prices. Working Paper Series.
Gargano, A. and Timmermann, A. (2012). Predictive Dynamics in Commodity Prices. Danish National Research Foundation.
Husain, A. and Bowman, C. (2004). Forecasting Commodity Prices: Futures versus Judgment. IMF Working Papers, 04(41), p.1.
K, S. and Sasithra, S. (2014). REVIEW ON CLASSIFICATION BASED ON ARTIFICIAL NEURAL NETWORKS. International Journal of Ambient Systems and Applications (IJASA), 2(4).
Khalid, M., Sultana, M. and Zaidi, F. (2014). Prediction of Agriculture Commodities Price Returns Using ARMA and Wavelet. Journal of Natural Sciences Research, 4(23).
Kulkarni, S. and Haidar, I. (2009). Forecasting Model for Crude Oil Price Using Artificial Neural Networks and Commodity Futures Prices. International Journal of Computer Science and Information Security, 2(1).
Liu, C., Hoi, S., Zhao, P. and Sun, J. (2016). Online ARIMA Algorithms for Time Series Prediction. roceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16).
Wets, R. and Rios, I. (2015). Modeling and estimating commodity prices: copper prices. Mathematics and Financial Economics, 9(4), pp.247-270.
Essay Writing Service Features
Our Experience
No matter how complex your assignment is, we can find the right professional for your specific task. Contact Essay is an essay writing company that hires only the smartest minds to help you with your projects. Our expertise allows us to provide students with high-quality academic writing, editing & proofreading services.
Free Features
Free revision policy
$10Free bibliography & reference
$8Free title page
$8Free formatting
$8How Our Essay Writing Service Works
First, you will need to complete an order form. It's not difficult but, in case there is anything you find not to be clear, you may always call us so that we can guide you through it. On the order form, you will need to include some basic information concerning your order: subject, topic, number of pages, etc. We also encourage our clients to upload any relevant information or sources that will help.
Complete the order form
Once we have all the information and instructions that we need, we select the most suitable writer for your assignment. While everything seems to be clear, the writer, who has complete knowledge of the subject, may need clarification from you. It is at that point that you would receive a call or email from us.
Writer’s assignment
As soon as the writer has finished, it will be delivered both to the website and to your email address so that you will not miss it. If your deadline is close at hand, we will place a call to you to make sure that you receive the paper on time.
Completing the order and download