CN110059852A - A kind of stock yield prediction technique based on improvement random forests algorithm - Google Patents

Abstract

The invention discloses a stock return rate prediction method based on an improved random forest algorithm. The present invention aims at the difficulty of parameter selection and classification performance problems in the random forest when the stock return rate is classified and predicted, and the RF algorithm itself cannot identify and select a more efficient one. The shortcomings of the features, combined with the particle swarm algorithm to optimize the feature selection mechanism, when the trend change is not obvious in the early stage, the optimal features are screened out, and they are input into the RF algorithm as attributes, and a hybrid method of PSO-GRID-RF stock trend prediction is proposed; this paper The invention reduces feature subsets, eliminates irrelevant or duplicated feature attributes, reduces the dimension of input, and reduces the time for stock trend prediction; in the multi-attribute feature environment, an efficient feature selection method is proposed, and a grid search algorithm is introduced to optimize Random forest parameters, thereby improving the classification prediction performance of random forest, and greatly improving the accuracy of stock trend prediction.

Description

A Stock Return Prediction Method Based on Improved Random Forest Algorithm

technical field

The invention belongs to the technical field of financial data mining. Aiming at the difficulty of parameter selection and the problem of classification performance in random forests in the classification and prediction research of stock returns, a particle swarm algorithm-based feature selection and grid search algorithm are proposed to optimize random forests. New algorithm for parameters. The particle swarm algorithm is used to select the features of the training set, and the redundant indicators in the index system are eliminated to reduce the input dimension. At the same time, the grid search algorithm is introduced to optimize the parameters of the random forest, thereby improving the classification and prediction performance of the random forest.

Background technique

In the stock market, prediction of stock price movements has always been a hot issue for investors. Accurately judging and grasping the changing trend of the entire stock market can not only reduce the phenomenon of blind investment in the stock market, but also has a high practical significance for improving the rationality of investors in the stock market, and can also provide a reference for the country to formulate relevant economic policies.

Scholars at home and abroad have conducted in-depth research on stock price forecasting and have proposed various forecasting methods. There are two main methods currently used, fundamental analysis and technical analysis. The first category is based on consideration of fundamental factors such as company growth and profitability. The second type is mathematical analysis based on past stock data, this simple analysis is to make predictions by observing the trend chart of stock movements. More sophisticated analyses employ sophisticated statistical methods and machine learning algorithms.

Time series analysis is the first method applied to stock price forecasting, which establishes an ARMA model for short-term forecasting of stock opening prices. Due to the influence of various factors, the stock price presents nonlinear changes. The time series analysis method based on the linear model cannot reflect the nonlinear change law of the stock well, and the prediction accuracy is low, and the application is limited. With the rise of artificial intelligence technology, BP neural network is widely used in stock price prediction because of its powerful nonlinear mapping ability. Based on the stock price prediction model SPPM of BP neural network, multiple neural network models are established to predict stock prices. .Neural network has achieved good results in nonlinear stock prediction, but at the same time there are problems of unstable learning and memory, slow convergence speed, and easy to fall into the local optimal value.

Random Forest algorithm (Random Forest) has been applied in the financial field as a classification technology. Compared with Support Vector Machine (Support Vector Machine) and Artificial Neural Networks (Artificial Neural Networks), RF is better in stock trend prediction the result of. The random forest algorithm is a combination of models, and it has achieved good results in different fields. Based on the advantages of the random forest algorithm, such as fast training speed and strong model generalization ability, the application of this algorithm to the prediction of stock fluctuations can avoid the shortcomings of the above prediction models. Random forest prediction is mainly to first screen the established initial indicator system, and then substitute the screened indicator data into the random forest as influence variables, and output the fluctuations as response variables. However, the existing methods lack the model optimization of the random forest itself, and cannot further improve the prediction accuracy.

SUMMARY OF THE INVENTION

Aiming at the technical deficiencies in the research on the classification and prediction of stock returns, the invention proposes a stock return prediction method based on an improved random forest algorithm.

A stock return prediction method based on an improved random forest algorithm, which specifically includes the following steps:

Step 1: Data acquisition, get the stock daily data through the website;

Divide the data into training set, validation set, test set

Step 2: Get the data for exponential smoothing:

S ₀ =Y ₀ t=0 (1)

S _t =α*Y _t +(1-α)*S _t-1 t>0 (2)

In the formula: S _t represents the smoothed value of time t, Y _t represents the actual value of time t;

S ₀ represents the smoothed value of the data at t=0, Y ₀ represents the actual value at t=0, and t represents the number of days on which the stock was acquired;

S _t-1 time is the smoothing value of t-1, α is the exponential smoothing factor, 0<α<1.

Exponential smoothing removes randomness or noise from changes in historical data, allowing models to easily identify long-term price trends.

Step 3: Feature extraction: Calculate the technical indicators according to the exponential smoothing results, calculate the feature matrix from the smoothed time series data, and use the technical indicators used by investors to judge the stock trend as a feature.

Step 4: PSO algorithm for feature selection:

To determine the necessary impact indicators as the input of the model, and the necessary response variables as the output of the model, because the construction of a stock indicator system is the basis for subsequent evaluation and comprehensive analysis, we perform feature extraction on technical indicators. We take the technical index as the particle in the particle swarm algorithm. The initial velocity and position of the particle in the PSO algorithm are randomly assigned, the local optimal solution P _idbest is the optimal position of the particle in the current iteration, and the global optimal solution P _gdbest is the optimal location for the entire population. Assuming that the dimension of the particle swarm search space is D, and there are m particles in total, the position of the particle in the space is x _i =[x _i1 ,x _i2 ,...,x _iD ], and the speed is _{vi =[v i1 ,} v _i2 ,... ,v _iD ], i=1...m, the calculation formula is as follows:

Adjust space position

In the formula: V ^k corresponds to the velocity of the local extremum of a particle in the k-th dimension, X ^k is the optimal position of the k-th dimension of the local value of the particle, represents the local optimal position of the particle swarm in the k-th iteration process, Represents the global optimal position of the particle swarm in the k-th iteration process, S( ) represents the sigmoid function, and takes the velocity as the variable of the sigmoid function. To adjust the spatial position, the particle velocity is mapped to between [0, 1] and combined with Random number comparison, update the position state of the particle, c ₁ , c ₂ are learning factors, and are positive numbers, w is the inertia weight, rand ₁ , rand ₂ ∈ [0,1], random uniform distribution;

Step 5: Set the judgment conditions:

Set the judgment condition: if the number of iterations exceeds the maximum number of iterations and the fitness is lower than the set value, the loop will be jumped out.

Step 6: Feature Selection:

The binary code obtained by the particle swarm feature selection in step 4 is used as the input feature for trend prediction, where 1 means selected and 0 means not selected;

Step 7: Output optimal features:

If step 5 is satisfied, the set condition outputs the optimal feature, otherwise returns to step 4;

Step 8: Build the data matrix:

Build a data matrix according to the optimal feature selected in step 7;

Step 9: Training set, validation set cross-validation:

The training set and the validation set are adjusted by cross-validation, 90% are used to train the model and 10% are used to validate the model. Use the grid search algorithm to optimize the parameters of the random forest, including the depth of the tree, the random state, the number of variables in the tree node, the number of trees, the OOB misclassification rate, and the estimation of variable importance to improve the prediction accuracy, so as to obtain the prediction. model, so that the model has better fitness to the data.

Step 10: Establishment of stock trading signals, namely data labels:

Use the data matrix constructed in step 8 as training data, input it into the random forest algorithm for training, and construct the trading signal Yj={y1,y2,...,yj}, where j=1,2,...,n is the sample number. The specific construction steps of the trading signal are as follows:

1) Calculate the daily average price p _j

Among them, C _j represents the closing price of the stock, H _j represents the highest price of the stock, and L _j represents the lowest price.

2) Calculate the arithmetic return V _j of the next k days, k=1, 2, . . . , 10;

3) Construct the trading signal y _j

Step 11: In-Sample Training:

Input the data matrix constructed by the optimal features into the random forest algorithm model for training, and use the grid search algorithm to optimize the parameters of the new optimal feature data set, and compare it with the actual stock trend to obtain the stock forecast trend and forecast. accuracy.

Step 12: Model Evaluation:

In the classification process according to the random forest algorithm, the classification prediction results can be represented by a confusion matrix, as shown in Table 1 below:

Table 1 Confusion Matrix

Predicted +1 Predicted to be 0 Predicted as -1 True is +1 TP FZ1 FN1 true is 0 FP1 TZ FN2 true is -1 FP2 FZ2 TN

Among them, TP is +1 for correct classification, 0 for correct classification in TZ, TN for negative class for correct classification, FP1 for class 0 error is divided into +1 class, FP2 for -1 class error is divided into +1 class, FZ1 is +1 class Class error is divided into 0 class, FZ2 is -1 class error is classified into 0 class, FN1 is +1 class error is classified into -1 class, FN2 is 0 class error is classified into -1 class, FP is FP1+FP2, FN (is FN1+FN2, FZ is FZ1+FZ2, and N=NTP+NFN+NFP+NTN+NTZ represents the total amount of samples.

Accuracy represents the probability of being correctly predicted in the test set, Recall represents the probability that the samples of the positive class were predicted to be correct, and Precision represents the correct probability of all samples that were predicted to be positive classes. The calculation formulas are as follows: :

Recall=TP/(TP+FN) (10)

Precision=TP/(TP+FP) (11)

F is a comprehensive performance index composed of the weighted average of sensitivity and precision. The closer the F value is to 1, the better the classification result. The formula is as follows:

The above parameters are obtained from the chaos matrix. On the other hand, in the random forest generation process, the training set is generated by the bootstrap method. Due to the repeated sampling with replacement, only about 63% of the data is generated compared to the original data. Repeat the extraction, and the rest of the data will not appear, of which the rest of the data is the out-of-bag data OOB. The out-of-bag data is used to estimate the generalization ability of the random forest algorithm, which is called OOB estimation; with a tree as the unit, the OOB data is used to detect The correct rate is OOB _score , the detected error is the out-of-bag error OOB _error , and the average OOB _error of all trees is the OOB' _error of the random forest. The smaller the OOB' _error , the stronger the generalization ability of RF; the fitness The value Fitness is composed of F and OOB' _error . The smaller the value, the better. The formula is as follows:

OOB _error = 1-OOB _score (13)

Fitness=OOB' _error +(1-F) (14)

Step 13: Out-of-sample testing:

After determining the optimal parameters, use the test data to test the trained random forest algorithm model to obtain the classification results, and use all the preprocessed sample features of the test set as the input of the model to obtain the T+k prediction value of each sample. To the classification results, and compared with the actual stock trend, the trend of stock forecast and the accuracy of the forecast.

The beneficial effects of the present invention are as follows:

(1) The present invention proposes an efficient feature selection method. Through the global search of the PSO algorithm, the best feature is selected as an input variable and input to the RF algorithm, which reduces the feature subset, eliminates irrelevant or duplicated feature attributes, and reduces the number of stocks. The prediction time greatly improves the accuracy of stock trend prediction.

(2) The present invention adopts cross-validation in the training set, effectively considers the time-series correlation of stock prices, and effectively improves the accuracy of the random forest classification model.

(3) The stock return rate is a stationary sequence. Using the stock return rate as a label can better reflect the price trend than using the closing price as an input label, which can effectively improve the accuracy of stock trend prediction.

(4) The present invention adopts the grid search algorithm to optimize the parameters in the parameter training of the random forest, which effectively avoids the difficult problem of parameter selection when the random forest algorithm performs prediction, selects the best parameters, and improves the accuracy of trend prediction.

Description of drawings

Figure 1 is a framework diagram of the stock return research of improved random forest;

Figure 2 is an undirected graph of random forest classification;

Fig. 3 is a schematic diagram of binary encoding;

Fig. 4 is the PSO algorithm flow chart;

Figure 5 is a flowchart of the random forest voting classification method in stock trend prediction.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

As shown in Figures 1-5, the present invention is based on particle swarm algorithm, grid search algorithm, random forest algorithm, and a stock trend prediction method for stock return research.

The present invention proposes a stock return research method of an efficient feature selection method in a multi-attribute feature environment. Since the attribute features are large in scale and all belong to continuous variables, the CART algorithm is used to generate a decision tree, and the specific formula is as follows:

The flow chart of the present invention is shown in Figure 1, and the concrete steps are as follows:

(1): Data acquisition:

The stock daily data is obtained through the website. The stock data source used in the present invention is a website such as Yahoo (Yahoo), including the opening price, closing price, trading volume, highest price, lowest price, etc. of stock trading, downloaded as a CSV file, and the The data is divided into training set, validation set, and test set.

(2): Obtain data for exponential smoothing:

S ₀ =Y ₀ t=0 (3)

S _t =α*Y _t +(1-α)*S _t-1 t>0 (4)

where S _t represents the smoothed value at time t, and Y _t represents the actual value at time t;

S ₀ represents the smoothed value of the data at t=0, Y ₀ represents the actual value at t=0, and t represents the number of days on which the stock was acquired;

S _t-1 time is the smoothing value of t-1, α is the exponential smoothing factor, 0<α<1.

Exponential smoothing removes randomness or noise from changes in historical data, allowing models to easily identify long-term price trends.

(3): Feature extraction;

Technical indicators are calculated based on the results of exponential smoothing. The indicators should consider all aspects of market behavior. The specific values and interrelationships of indicator values directly reflect the state of the stock market and provide guidance for our operational behavior. Most of the indicators reflect things that are not directly visible from the information sheet. The technical indicators calculate the feature matrix for the smoothed time series data, and use the technical indicators accepted by investors to judge the ups and downs of the stock trend as features.

(4): PSO algorithm for feature selection:

Figure 4 is a flowchart of the PSO algorithm, the steps of feature selection by particle swarm. In the PSO algorithm, the optimization problem is transformed into a point in the d-dimensional space, called a particle. The current position of the particle is evaluated by the objective function, and the objective function calculates the corresponding fitness according to the position of the particle. The particle flies in the search space at a certain speed, which is dynamically adjusted according to its own flying experience and the flying experience of its companions, and is then used to calculate the new position of the particle. The optimization search is performed in an iterative manner in a population composed of a group of randomly initialized particles until a certain termination condition is met, such as a specified number of iterations.

In the PSO algorithm, the initial velocity and position of the particles are randomly assigned, the local optimal solution P _idbest is the optimal position of the particle in the current iteration, and the global optimal solution P _gdbest is the optimal position of the entire population. Assuming that the dimension of the particle swarm search space is D, and there are m particles in total, the position of the particle in the space is x _i =[x _i1 ,x _i2 ,...,x _iD ], and the speed is _{vi =[v i1 ,} v _i2 ,... ,v _iD ], i=1...m, the calculation formula is as follows:

Adjust space position

In the formula: corresponding to the velocity of the local extreme value of a particle in the kth dimension, the optimal position of the particle local value in the kth dimension represents the local optimal position of the particle swarm in the k-th iteration process, and represents the particle swarm in the k-th iteration process. The global optimal position of , S( ) represents the sigmoid function, and the velocity is used as the variable of the sigmoid function. To adjust the spatial position is to map the particle velocity to between [0, 1] and compare it with a random number to update the position state of the particle , is the learning factor, and is a positive number, is the inertia weight, , randomly and uniformly distributed;

(5): Set the judgment conditions:

Set the judgment condition: The condition judgment is shown in Figure 4. If the number of iterations exceeds the maximum number of iterations and the fitness is less than the set value, the loop will be jumped out.

(6): Feature selection:

The binary code defines whether a feature is selected as an input feature for trend prediction, as shown in Figure 3, where 1 means selected and 0 means not selected;

(7): Output optimal features:

If (5) is satisfied, the set condition will output the optimal feature, otherwise, return to (4);

(8): Build a data matrix:

Construct a data matrix according to the optimal feature selected in (7);

(9): The training set and the cross-validation set are combined:

The training set and the cross-validation set are combined: the training set is adjusted by cross-validation, 90% is used for training the model, and 10% is used for validating the model. Use the grid search algorithm to optimize the parameters of the random forest, including the depth of the tree, the random state, the number of variables in the tree node, the number of trees, the OOB misclassification rate, and the estimation of variable importance to improve the prediction accuracy, so as to get The prediction model makes the model have better fitness for the data.

(10): Establishment of stock trading signals:

Using the data matrix constructed in (8) as training data, input it into the random forest algorithm for training, and construct the trading signal Yj={y1,y2,...,yj}, where j=1,2,...,n is the sample number. The specific construction steps of the trading signal are as follows:

1) Calculate the daily average price p _j

Among them, C _j represents the closing price of the stock, H _j represents the highest price of the stock, and L _j represents the lowest price.

2) Calculate the arithmetic return V _j of the next k days, k=1, 2, . . . , 10;

3) Construct the trading signal y _j

(11): Random forest for classification prediction:

In life, our cognition of things is based on the judgment and classification of characteristics, for example, mammals can be judged by whether they were born or not. Random forest adopts this idea, as shown in Figure 2, which is an undirected graph of random forest classification. At each node of the tree, the leaf nodes of the next layer are split by certain rules according to the performance of the feature, and the leaf node of the terminal is the final classification result. The key to random forest learning is to choose the optimal partitioning attribute. With the layer-by-layer division, the sample categories contained in the branch nodes of the decision tree will gradually converge, that is, when the nodes are split, the information gain after node splitting should be maximized.

Random Forest builds each decision tree according to the following two-step method. Specifically, the first step is called "row sampling", sampling with replacement from the entire training sample to obtain a Bootstrap dataset. The second step is called "column sampling", randomly select m features from all M features (m is less than M), and train a decision tree with the m features of the Bootstrap dataset as a new training set. The classification prediction is the category or one of the categories with the most votes cast from the N decision trees is the final category, as shown in the flowchart of the random forest voting classification method in the stock trend prediction in Figure 5. Random forest model construction can achieve the effect of reducing the probability of overfitting. In the random forest, although each tree is divided by only m factor features, the classification effect is not good alone, but it is more stable when combined. It may be understood in this way that each decision tree is an expert who is proficient in a certain narrow field (select m from M factors for each tree to learn), and a random forest contains many experts who are proficient in different fields. The problem (new data set), you can look at it from different perspectives, and finally vote to get the result.

(12): Principle of grid search algorithm:

Grid search is an exhaustive search method that specifies parameter values, then each combination is used for random forest training and cross-validation is used to evaluate performance. After the fit function has tried all parameter combinations, it returns a suitable classifier and automatically adjusts to the best parameter combination.

(13): In-sample training:

Input the data matrix constructed by the optimal features into the random forest algorithm model for training, and use the grid search algorithm to optimize the parameters of the new optimal feature data set, and compare it with the actual stock trend to obtain the stock forecast trend and forecast. accuracy.

(14): Out-of-sample testing:

After determining the optimal parameters, use the test data to test the trained random forest algorithm model to obtain the classification results, and use all the preprocessed sample features of the test set as the input of the model to obtain the T+k prediction value of each sample. To the classification results, and compared with the actual stock trend, the trend of stock forecast and the accuracy of the forecast.

Claims (1)

1. a stock rate of return prediction method based on improved random forest algorithm, is characterized in that, specifically comprises the following steps: Step 1: Data acquisition, get the stock daily data through the website; Divide the data into training set, validation set, and test set; Step 2: Get the data for exponential smoothing: S ₀ =Y ₀ t=0 (1) S _t =α*Y _t +(1-α)*S _t-1 t>0 (2) where S _t represents the smoothed value at time t, and Y _t represents the actual value at time t; S ₀ represents the smoothed value of the data at t=0, Y ₀ represents the actual value at t=0, t represents the number of days to obtain the stock day; S _t-1 time is the smoothed value of t-1, α is the exponential smoothing factor, 0 <α<1; Step 3: Feature Extraction Calculate technical indicators according to the results of exponential smoothing, and calculate the characteristic matrix of smoothed time series data, and use the technical indicators used by investors to judge stock trends as features; Step 4: PSO algorithm for feature selection Taking the technical indicators as the particles in the particle swarm algorithm, the initial speed and position of the particles in the PSO algorithm are randomly assigned, the local optimal solution P _idbest is the optimal position of the particle in the current iteration, and the global optimal solution P _gdbest is the entire population. Optimal position; Assuming that the dimension of the particle swarm search space is D, and there are m particles in total, the position of the particle in the space is x _i =[x _i1 ,x _i2 ,...,x _iD ], and the speed is v _i =[v _i1 , v _i2 ,...,v _iD ], i=1...m, the calculation formula is as follows: Adjust space position In the formula: V ^k corresponds to the velocity of the local extremum of a particle in the k-th dimension, X ^k is the optimal position of the k-th dimension of the local value of the particle, represents the local optimal position of the particle swarm in the k-th iteration process, Represents the global optimal position of the particle swarm in the k-th iteration process, S( ) represents the sigmoid function, and takes the velocity as the variable of the sigmoid function. To adjust the spatial position, the particle velocity is mapped to between [0, 1] and combined with Random number comparison, update the position state of the particle, c ₁ , c ₂ are learning factors, and are positive numbers, w is the inertia weight, rand ₁ , rand ₂ ∈ [0,1], random uniform distribution; Step 5: Set the judgment conditions: Set the judgment condition: if the number of iterations exceeds the maximum number of iterations and the fitness is lower than the set value, the loop will be jumped out; Step 6: Feature Selection: The binary code obtained by the particle swarm feature selection in step 4 is used as the input feature for trend prediction, where 1 means selected and 0 means not selected; Step 7: Output optimal features: If step 5 is satisfied, the set condition outputs the optimal feature, otherwise returns to step 4; Step 8: Build the data matrix: According to the optimal feature selected in step 7, construct the data matrix of the input random forest; Step 9: Cross-validation on training set and validation set: In order to improve the prediction accuracy of random forest, the training set and validation set are adjusted by cross-validation. The depth of the tree, the random state, the number of variables in the tree node, the number of trees, the OOB misclassification rate, and the estimation of the importance of variables are used to improve the prediction accuracy, so as to obtain a prediction model, which makes the model have better fitness for the data. high precision; Step 10: Establishment of stock trading signals, namely data labels: Use the data matrix constructed in step 8 as training data, input it into the random forest algorithm for training, and construct the trading signal Yj={y1,y2,...,yj}, where j=1,2,...,n is the sample number; the trading signal The specific construction steps are as follows: 1) Calculate the daily average price p _j Among them, C _j represents the closing price of the stock, H _j represents the highest price of the stock, and L _j represents the lowest price; 2) Calculate the arithmetic return V _j of the next k days, k=1, 2, . . . , 10; 3) Construct the trading signal y _j Step 11: In-Sample Training: Input the data matrix constructed by the optimal features into the random forest algorithm model for training, and use the grid search algorithm to optimize the parameters of the new optimal feature data set, and compare it with the actual stock trend to obtain the stock forecast trend and forecast. accuracy; Step 12: Model Evaluation: In the classification process according to the random forest algorithm, the classification prediction results can be represented by a confusion matrix, as shown in Table 1 below: Table 1 Confusion Matrix Predicted +1 Predicted to be 0 Predicted as -1 True is +1 TP FZ1 FN1 true is 0 FP1 TZ FN2 true is -1 FP2 FZ2 TN Among them, TP is +1 for correct classification, 0 for TZ correct classification, TN for negative class for correct classification, FP1 for class 0 error is divided into +1 class, FP2 for -1 class error is divided into +1 class, FZ1 is +1 class Class error is divided into 0 class, FZ2 is -1 class error is classified into 0 class, FN1 is +1 class error is classified into -1 class, FN2 is 0 class error is classified into -1 class, FP is FP1+FP2, FN (is FN1+FN2, FZ is FZ1+FZ2, N=NTP+NFN+NFP+NTN+NTZ represents the total amount of samples; Accuracy represents the probability of being correctly predicted in the test set, Recall represents the probability that the samples of the positive class were predicted to be correct, and Precision represents the correct probability of all samples that were predicted to be positive classes. The calculation formulas are as follows: : Recall=TP/(TP+FN) (11) Precision=TP/(TP+FP) (12) F is a comprehensive performance index composed of the weighted average of sensitivity and precision. The closer the F value is to 1, the better the classification result. The formula is as follows: The above parameters are obtained from the chaos matrix. On the other hand, in the random forest generation process, the training set is generated by the bootstrap method. Due to the repeated sampling with replacement, only about 63% of the data is generated compared to the original data. Repeat the extraction, and the rest of the data will not appear, of which the rest of the data is the out-of-bag data OOB. The out-of-bag data is used to estimate the generalization ability of the random forest algorithm, which is called OOB estimation; with a tree as the unit, the OOB data is used to detect The correct rate is OOB _score , the detected error is the out-of-bag error OOB _error , and the average OOB _error of all trees is the OOB' _error of the random forest. The smaller the OOB' _error , the stronger the generalization ability of RF; the fitness The value Fitness is composed of F and OOB' _error . The smaller the value, the better. The formula is as follows: OOB _error = 1-OOB _score (14) Fitness=OOB' _error +(1-F) (15) Step 13: Out-of-sample testing: After determining the optimal parameters, use the test data to test the trained random forest algorithm model to obtain the classification results, and use all the preprocessed sample features of the test set as the input of the model to obtain the T+k prediction value of each sample. To the classification results, and compared with the actual stock trend, the trend of stock forecast and the accuracy of the forecast.