Financial derivative features based integrated potential fishing zone (IPFZ) Future forecast

3.1Dataset

Ocean Surface Temperature data (OSTs), chlorophyll, and Ocean Sat II - India (Fig. 2) Optical Communication Band, and MODIS Aqua (US) are used to detect potential fishing zones all over the place of the Indian coastline, which are regularly together by NOAA-AVHRR thermal-infrared frequencies and Eumetsat (ESA) Met-OP satellites. For each sector PFZ Announcements are produced in the form of PFZ Maps and Texts, and where PFZ are positioned (latitude and longitude data). Because of the sea’s lively atmosphere, the selected place could be progressed by recognised fishing zones on the atlases. The wind speediness and route information is also incorporated into the PFZ maps to guide fisherman on the likely shifts in PFZ. This fact allows the fishermen to trace the PFZ in the maps; on the other hand after a day they will reach the site. The PFZ maps and text for the accessibility of fisherman are also obtained in each sector’s with their local languages. The PFZ text gives data about exact position of (length, latitude), depth at the PFZ site and the distance from important places on the coastlines (fishing landing stage centre, light erections) clearly identifiable.

Fig. 2

Retrieved Oceansat-2 Satellite Data Chlorophyl Image.

3.2Pre-processing & normalization

The unused data and the missing datas are cleaned using pre-processing methods. In addition, the shape of the data are altered during normalization method. In the beginning, the total number of the input uncooked data is around 151031 with 16 features. The whole raw datas are taken as training data and Table 1 describes the total number of features.

Among the 16 features UPDATED_DATE is use as targeted data to training the suggested model. The input information is taken from 15-5-2013 to 18-7-2018, where the main aim of the objective is to predict the potential fishery zone by giving future data. Hence, only 15 unique features are used in this research, however, only 13 unique features are used for training the system. The reason is that the feature LONGITUDE_DIRECTION (East) and LATITUDE_DIRECTION (North) has same information for 151031 data, which is not useful for training the proposed model. In order to training the projected network and attain improved classification precision, the “-“ between the date is removed for the period of pre-processing. The DIRECTION feature has seven directions as Southern East (SE), Southern West (SW), Northern East (NE), Southern (S), East (E), Northern (N) and Western (W). This datas are cannot be use openly to train the projected model, hence these characters are rehabilitated into variable. In addition, the Landing_Center attribute are in an floating format, which is altered into double layout for improved prediction. Hence, 13 features are known as input to the FFNN for predicting the PFZ. Moreover, in order to accurately predict the future PFZ, this research study uses more quantity of features such as Economic Derivatives and geometrical metrics, which is described as follows.

3.3Financial derivative features

The way the dataset are deal with is another deliberation, as it is uncommon for raw data values to be give back to an algorithm to be used. The data should consequently be revised to fit our challenging domain. The aim of this research work is to predict the IPFZ over the well-defined region based on the collected fish capacity. An important factor that needs to be considered is the need to come to an end of future agreements, usually up to one year in advance, and the indenture period might take any length. In addition, given the unique character of the PFZ projections, there is an even higher need for data transformation. Therefore, we employ an Economic Derivative feature (Table 2) to give the data more understandable and more appropriate for the problem area. The result looks to be much less altered, which is why financial derivatives are applied, with adding, volatility on a daily basis appears to be lesser and a pattern is ease to recognize in PFZ predicting. This approach of predicting IPFZ is extremely flexible and may be altered to any length of interest.

3.4Classification

The more number of features are set as input to the proposed FFNN technique for forecasting the PFZ, where the different types of neural networks are used to validate the presentation of the projected ideal. These neural networks include ENN, RNN and CFN are also discussed as in Comparative Techniques.

3.4.1Proposed Function Fitting Neural Network (FFNN)

A neural network associates a number of levels of processing, engaging basic, simultaneous and biologically nervous-inspired parts. It encompasses of a layer of input, one or numerous hidden layers and a layer of output. In every layer, the input of each layer involves of a number of nodes, or neurons, so that neurons join the different layers. Typically every neuron has weights modified during the learning process and modifies the strength of the signal of this neuron as the weight decreases or grows. The structure of a FFNN is defined in the Fig. 3.

Fig. 3

Single node in a FFNN network.

The inputs of the neuron x_k, k = 1, … k, and the continuous bias term θ_i are multiplying and added up. The result of n i shall be the g-enabled input. Initially, however, a hyperbolic tangent (tanh) or a sigmoid function is the most common activation function for the mathematical accessibility. Tangent is defined as hyperbolic

(1)

tanh(x)=1-e-x1+e-x

The node i becomes

(2)

yi=gi=g(∑j=1Kwji+xj+θi)

Linking many nodes in parallel and series, a FFNN network is intended. A typical network for single hidden layer is shown in Fig. 4.

Fig. 4

FFNN with one hidden layer.

The output, y_i i = 1 and 2, of the FFNN network becomes

(3)

yi=g(∑j=13wji2g(nj1)+θj2)   =g(∑j=13wji2g(∑k=1Kwkj1+θj1)+θj2)

From (3) we may deem a non-linear map from the input space x to the output space (m = 3), the FFNn is a non-linear parameterised map (here m = 3). The weights of Wji and biases are parameters of θjk . In general, the g active functions are presumed to be the same and known in every layer beforehand. The g function is used for all layers in the figure.

When you use data (x_i, y_i) i = 1, … . N, it is a data fitting issue to discover the optimal FFNN network. The parameters (wjik,θjk) to be determined are.

This is the procedure. The designer must first f the architecture of the FFNN network: the number of layers and neurons concealed in each layer. Figure 5 shows the total architecture. At this point, too, the activation functions for each layer are expected to be known. The weights and biases (wjik,θjk) are unknown parameters to be calculated.

Fig. 5

The Proposed MATLAB FFNN Architecture.

After procurement the results from FFNN, the post-processing is passed out. In this strategy, the twofold organize of LC include is changed over into characters. The factors of DIR are adjusted into seven directions. Subsequently, the comes about will be in an appropriate arrange, for occasion in case an client allow a future information as input, the location label name, its flow direction, latitude and longitude bearings are shown as yield of FFNN show [15].

3.5Comparison neural network models

3.5.1Elman neural network

A feedback neural system improved by Elman in 1990 can be characterized as the Elman Neural Network (ENN). ENN focuses on the training of the neural background network (BPNN). The Elman neural network architecture is usually divided into 4 layers. The undertaking layer function is to save the output of the cached layer. The output of the hidden layer connections its entrance through the delay and the storage of the undertaking layer [16] since it is focused on the background of a neural network. ENN Architecture is illustrated in Fig. 6.

Fig. 6

Architecture of ENN.

The output of the hidden components of its input by delaying and storing the company layer are based on the BP network. This joining is sensitive to past information and the network for interior input can increase the dynamic data management capacity. The internal state memorization allows for a dynamic planning method, which allows the system to adapt to temporal alterations across interval. [17].

The weight of the hidden layer input is w 1, the weight of the hidden layer in the layer is w₂, the weight of the hidden layer to the output layer is w₃; u (k - 1) is the input for the neural network; x (k) is the output for the hidden layer; y (k) is an output for a hidden layer, x_c (k) is the output of the undertaking layer and y(k) is the output of that hidden layer. x_c (k) is the output of the undertaking layer and

(4)

x(k)=f(w2xc(k)+w1(u(k-1)))

Where

(5)

xc(k)=x(k-1)

F is the transfer function hidden layer, which is usually used in S-type function; that is:

(6)

f(x)=(1+e-x)-1

g is the output layer transfer function, which is often a linear function; that is:

(7)

y(k)=g(w2x(k))

To revise weights and the error of network is:

(8)

E=∑k=1m(tk-yk)2

where t_k is represent as the output vector of the object.

5.2Recurrent Neural Network (RNN)

Because of its extra feedback from the created layer output between input and output layers, RNN is an exclusive class of neural network.his layer is the context layer that retains data between observations [18]. In the present period, the results of the processing in the previous phase can be carried over and employed. Iparticular in real time applications, the fundamental characteristic of RNN offers an extremely large advantage. In adding to learning via all current inputs, RNN can have an unbounded memory level, and can thus learn connections over time [19]. The RNN is showed in Fig. 7.

Fig. 7

RNN Architecture.

Hidden layer of Input is stated as in (9) as:

Here h_t is represented as the hidden layer at the instance t^th, X_t is represented as the input at instance t^th, moreover, g_n is represented as the function, and W_xh is represented as the input to hidden layer of weight matrix Equation (10) which characterizes the concealed to output layer is stated as:

3.5.3Cascade forward network

The Cascade Forward Network (CFN), which uses back propagation to update weights, has similarities with the FFNN. The primary dtinction ithat the connection between the individual level and the input at the next level is weighted [20]. In many circumstances, it has been thought that some cascade bpropagation can perform better than FFNN [21]. An important feature of this CFN is that each neuron layer is linked to the whole preceding neuron layer [22].The pictographic representation of CFN is in Fig. 8.

Fig. 8

Cascade FNN.

The CFN mathematical equation is specified as:

(11)

y=∑i=1nfi(wixi)+fo(∑j=1Kwjofjh(∑i=1nwjihxi))

Input layer activation function is provided in the output layer as a fⁱ, whereas input layer weight to output layer is wii . In the event of the combination of bias and an input layer, the activation function of each neuron within the covered up level is written as f^h, in order to express equation (11) as

(12)

y=∑i=1nfi(wixi)+fo  (wb+∑j=1kwjofh(wjb+∑i=1nwijhxi)

The next section will show the performance analysis of proposed model.

4.1Evaluation metrics

The challenge evaluation measurements are utilized tsurvey our method’s execution in division and classification. The assessment criteria for division incorporate sensitivity (SE), specificity (SP), accuracy (AC), and false measures. The presentation requirements are as follows:

Where tp,  tn,  fp and fn signify the true negative, false positive, number of a true positive TP, and false negative FN, N describes the total number of elements.

4.1Performance study of proposed FFNN model for 80% -20% of dataset

In this section, two different types of analysis are carried out by considereing with and without financial derivative features to validate the performance of proposed FFNN model with various neural networks. Initially, the available data is splits into 80% -20%, i.e. 80% of data is used for training procedure and remaining 20% of data is used for testing process. Table 3 provides the experimental value of different neural networks without financial derivative features for the analysis of 80% -20% in available dataset.

The above Table 3 validate the performance of projected model in terms of precision, specificity, recall, sensitivity and gmean. The sensitivity and recall are same for the all techniques, i.e. the ENN model achieved 1.000 of sensitivity and recall, where the proposed FFNN model achieved 1.000 of sensitivity and recall. In the specificity experiments, the ENN, RNN, CFNN and FFNN achieved 83%, 81%, 79% and 85%, which proves that the proposed FFNN achieved better performance than other neural networks. The ENN and RNN achieved nearly 91% of gmean, CFNN achieved 89% of gmean, but the proposed FFNN achieved 92% of gmean and high value of precision (i.e.0.0660). However, the existing neural network achieved only 0.0550 of precision value. From this study, the proposed FFNN model attained better presentation even without financial derivative features. Table 4 and Fig. 9 shows the performance of FFNN model in terms of accurateness and F-measure.

Fig. 9

Accuracy and F-measure Evaluation of FFNN Without Having Financial Derivative Features.

In the accuracy analysis of without having the financial derivative features, the ENN, RNN, CFNN and proposed FFNN achieved the 83%, 81%, 80% and 85%. As well as, these techniques achieved the f-measure of 0.11, 0.09, 0.09 and 0.12. From these analysis, it is clearly proved that without financial derivative features also, the proposed FFNN model achieved better performance for 80% -20% of available dataset. The next Table 5 shows the validation of proposed FFNN model with existing models by considering the financial derivative features for the 80% -20% of available dataset.

In the specificity analysis, the ENN, RNN, CFNN and FFNN achieved 86%, 88%, 85% and 90%, where the sensitivity and recall of all techniques are 1.000, when implementing with financial derivative features. In the precision experiments, the FFNN achieved high performance (i.e. 0.0994), where the ENN and CFNN achieved nearly 0.0670. The ENN, RNN, CFNN and FFNN achieved gmeans of 92%, 94%, 92% and 95% that shows that the projected FFNN model achieved better presentation than other techniques. However, the FFNN achieved less performance, when it is not implemented with Financial Derivative features. This proves that these added features plays a major role for predicting the future IPEZ accurately. Table 6 and Fig. 10 shows the experimental investigation of proposed FFNN model in terms of accuracy and F-measure for the 80% -20% of available dataset.

Fig. 10

Performance Evaluation of Proposed FFNN With Financial Derivative Features for 80% -20% of available dataset.

In the accuracy analysis, the ENN and CFNN achieved low accuracy (i.e. nearly 86%), RNN achieved 88% and proposed FFNN achieved 90%. The ENN, RNN, CFNN and FFNN achieved the f-measure of 0.12, 0.15, 0.12 and 0.18 that proves FFNN achieved better performance than other neural techniques. Without financial derivative features, the proposed FFNN achieved less performance, which is proven in Tables 3 and 4. The next section will explain the experimental analysis for 60% -40% of existing dataset.

4.2Performance study of proposed FFNN model for 60% -40% of dataset

Here, the performance of projected FFNN technique is validated with the 60% of training data and 40% of testing statistics. Table 7 shows the analysis of FFNN in terms of several metrics without considering the financial derivative features.

The above table validate the performance of proposed model in terms of s ensitiveness, specificity, precision, recall and gmean. The sensitivity and recall are same for the all techniques, i.e. the RNN, CFNN model achieved 1.000 of sensitivity and recall, where the proposed FFNN model achieved 1.000 of sensitivity and recall. In the specificity experiments, the ENN, RNN, CFNN and FFNN achieved 80%, 82%, 74% and 86%, which proves that the proposed FFNN achieved better performance than other neural networks. The ENN and RNN achieved nearly 90% of gmean, CFNN achieved 86% of gmean, but the proposed FFNN achieved 93% of gmean and high value of precision (i.e.0.0708). However, the existing neural network achieved only 0.0450 of precision value. From this analysis, the proposed FFNN model achieved improved performance with financial derivative features. Table 8 and Fig. 11 shows the performance of FFNN model in terms of accurateness and F-measure with additional financial derivative features.

Fig. 11

Evaluation of FFNN With Financial Derivative Features in terms of accurateness and F-measure.

In the accurateness analysis of with the financial derivative features, the ENN, RNN, CFNN and proposed FFNN achieved the 80%, 82%, 74% and 86%. As well as, these techniques achieved the f-measure of 0.09, 0.10, 0.07 and 0.13. From these analysis, it is clearly proved that with financial derivative features also, the proposed FFNN model achieved higher performance for 60% -40% of available dataset. The next section will presents the analysis of future data prediction using proposed FFNN with other neural networks.

4.3Performance analysis of future data prediction

By using the available datasets, the proposed FFNN tries to predict the future IPFZ accurately in terms of exactness, recollection, sensitivity, specificity and gmean. This analysis is taken for 80% of training data and 20% of testing data and the experimental values are given in Table 9.

The sensitivity and recall are same for the all techniques, i.e. the ENN, RNN, CFNN model achieved 1.000 of sensitivity and recall, where the proposed FFNN model achieved 1.000 of sensitivity and recall. In the specificity experiments, the ENN, RNN, CFNN and FFNN achieved 58%, 54%, 50% and 64%, which proves that the proposed FFNN achieved better performance than other neural networks. The ENN, RNN and CFNN achieved nearly 75% of gmean, but the proposed FFNN achieved 80% of gmean and high value of precision (i.e.0.0277) than other neural networks (i.e.0.022 of precision). However, the CFNN achieved only 0.0199 of precision value. From this analysis, the projected FFNN model attained better performance with financial derivative features. Table 10 and Fig. 12 shows the performance of FFNN model in terms of accurateness and F-measure with additional financial derivative features.

Fig. 12

Evaluation of proposed FFNN With Financial Derivative Features for future data.

In the accuracy analysis of with the financial derivative features, the ENN, RNN, CFNN and proposed FFNN achieved the 58%, 54%, 50% and 64%. As well as, these techniques achieved the f-measure of 0.04, 0.04, 0.03 and 0.05. From this analysis, it is clearly proved that with financial derivative features also, the proposed FFNN model achieved higher performance for future data. However, the performance is less when compared with the analysis of 80% -20% and 60% -40% of data. The reason is that the layers of FFNN model is needs to improve for the prediction of future data.

4.4Comparative analysis of Proposed FFNN with various Existing approach

The above table clearly proves that the proposed FFNN model achieved better accuracy than various existing techniques. For example, the data mining approach [23] and HMM [24] achieved nearly 83% of accuracy, where SVM [25], RF [26] achieved nearly 79% of accuracy. But, the FFNN model achieved 90.94% of accuracy due to the usage of additional financial derivative features as input for predicting the future IPFZ accurately [27]. However, the performance needs improvement by modifying the layers of FFNN.