Simulating the Linkages Between Economy and Armed Conflict in India With a Long Short-Term Memory Algorithm.

This article analyzes the linkages between the economy and armed conflict in India using annual frequency data for the period 1989-2016, the maximum time period for which consistent data are available for the country. An adequate set of economic indicators was established to fully reflect the economic condition. Long short-term memory (LSTM), which is a machine-learning algorithm for time series, was employed to simulate the relationship between the economy and armed conflict events. In addition, LSTM was applied to predict the trend of armed conflict with two strategies: multiyear predictions and yearly predictions. The results show that both strategies can adequately simulate the relationship between the economy and armed conflict, with all simulation accuracies above 90%. The accuracy of the yearly prediction is higher than that of the multiyear prediction. Theoretically, the future state and trend of armed conflict can be predicted with LSTM and future economic data if future economic data can be predicted.

INTRODUCTION

A variety of definitions of armed conflict are available in the empirical literature and databases, but the most frequently employed definition is a combination of three elements: (i) armed force, (ii) organized actor, and (iii) social harm. In this study, an armed conflict event refers to “an incident where armed force was used by an organized actor against another organized actor, or against civilians, resulting in at least 1 direct death at a specific location on a specific date.” Armed conflicts have a detrimental effect on the economic and social well‐being of people (Shahzad, Zakaria, Rehman, Ahmed, & Fida, 2016). In recent years, the number of armed conflicts has exhibited an upward trend of increasing conflicts from 2,372 in 1989 to 6,542 in 2016 with 1.956 million deaths according to the Uppsala Conflict Data Program (UCDP) (Fig. 1). Armed conflict has caused tremendous disaster to international security.

Fig. 1

Historical trend of armed conflict and GDP inflows in the world. Frequency of armed conflict from the UCDP (http://www.ucdp.uu.se/) and the GDP data from the world bank (https://data.worldbank.org/).

Armed conflict is the result of the interaction among multiple subsystems, including climate, natural resources, human security, and societal stability. Multiple pathways and forms of feedback exist among these subsystems (Scheffran, Brzoska, Kominek, Link, & Schilling, 2012). Each subsystem and its influence on armed conflict have historically received a considerable amount of attention from academics and researchers (Almer & Boes, 2012; Blakeslee & Fishman, 2013; Gizelis & Wooden, 2010; Kawsar, 2013; Lujala, 2010; Mares, 2013; Østby, Urdal, Tadjoeddin, Murshed, & Strand, 2011; Schleussner, Donges, Donner, & Schellnhuber, 2016; Sekhri & Storeygard, 2013). Some researchers believe that subsystems, e.g., resource scarcity and climate change, can produce economic changes that will increase the risk of armed conflict (Miguel, Satyanath, & Sergenti, 2004). Therefore, economic conditions are considered to have a crucial role in armed conflicts. Fig. 1 shows the historical trend of armed conflict and GDP inflows in the world. After the 1997 Asian financial crisis, when the GDP decreased, armed conflict substantially heightened. This phenomenon also appeared during the subprime mortgage crisis.

The empirical literature has shown that the economy proves to be more decisive than other factors in predicting the incidence of armed conflict. Raleigh and Urdal (2007) used logistic regression to analyze the causes of armed conflicts, including land degradation, freshwater scarcity, population density, and economic change. In this article, economic change is expressed in terms of the GDP. Their research indicated that the influence of economy on armed conflict is higher than other factors. Hauge and Ellingsen (1998) used the same method to study the influence of the economy on armed conflicts on a global scale; their study also applies the GDP as an indicator of the economy. Their research obtained the same conclusion as Raleigh: the contribution of the economy to the occurrence of armed conflict is greater than that to environmental scarcity. Empirics have also shown that a negative relationship exists between events of armed conflict and economic conditions. Sudula (2012) used the ratio of imports plus exports to GDP as an economic indicator of international trade dependence and explored the relationship between the trade dependence and the international armed conflicts among South Asian countries by bivariate linear regression. The results show a negative correlation between trade dependence and international armed conflicts. Lu and Thies (2010) built three multinomial logit models to examine how trade interdependence affects territorial, policy, and regime conflicts. They used the same economic indicator of trade interdependence that was applied by Sudula. The results show that interdependence of trade significantly reduces the incidence of these three types of conflicts. In addition, some researchers have focused on the relationship between the economy and terrorism, which is one type of armed conflict. Shahbaz, Shabbir, Malik, and Wolters (2013) used the AutoRegressive Distributed Lag (ARDL) bounds testing approach to simulate the relationship among inflation, economic growth, and terrorism in Pakistan. They discovered that an increase in inflation increases terrorist attacks, while economic growth is also a major contributor to terrorism. Additionally, the relationship between terrorism and capital was determined to be bidirectional (Shahbaz, 2013; Shahbaz et al., 2013). Malik and Zaman (2013) evaluated the short‐ and long‐term relationship between terrorism and economic factors, including price level, poverty, and political instability, via sophisticated econometric techniques including cointegration theory, the Granger causality test, and variance decomposition (VDC). The results show that economic conditions are drivers of terrorist attacks. Bukhari and Masih (2016) used the ARDL bounds testing approach followed by VDC and impulse response (IR) function to determine the causal relationship among the variables of economic growth, trade, and military spending on the onslaught of terrorism in Pakistan. They found that terrorism is most affected by the variables of trade and GDP in the short term. However, increased economic growth and military spending breed terrorism in the long term.

In general, the existing literature strongly demonstrates a quantitative relationship between the economy and armed conflicts. However, the following deficiencies exist in existing research: (1) The economic indicators are relatively solitary and cannot fully reflect the economic status of a country. (2) These studies relied on simple regression models or statistical models to mainly describe whether different economic indicators have positive or negative effects on armed conflict. Predicting armed conflicts is possible but very difficult, because armed conflict is a very complicated process. These models cannot capture the varying effects and complex interactions of the economy and armed conflict (Cederman & Weidmann, 2017).

Therefore, the main objectives of this study are as follows: (1) building an adequate set of economic indicators that can fully reflect the economic condition of a country; (2) simulating the relationship between the economy and armed conflict using a machine‐learning method that can reflect the complex interactions of the economy and armed conflict.

MATERIALS AND METHODS

Machine‐Learning Algorithm on Time Series

Regarding the prediction problem of the time series, conventional time series analysis methods or machine learning algorithms can be employed. For example, the autoregressive integrated moving average (ARIMA) and generalized autoregressive conditional heteroskedasticity (GARCH) are traditional methods for analysing time series that have successfully solved various time series prediction problems (Contreras, Espinola, Nogales, & Conejo, 2002; Garcia, Contreras, Akkeren, & Garcia, 2005). However, these algorithms use historical observations for deriving estimates of current and future values of the dependent variable that cannot be easily adapted to the prediction of multifactor inputs. Machine learning algorithms, such as support vector machines (SVMs) and the gray model (GM), are also applied to time series analysis (Kayacan, Ulutas, & Kaynak, 2010; Sapankevych & Sankar, 2009). These methods can introduce multiple factors to transform unsupervised learning problems into supervised learning problems. However, it cannot solve the problem of long‐term dependence in time series. There is a time lag in the impact of economic changes on armed conflicts. For example, this year's economic downturn may be the cause of the outbreak of armed conflict in next year. Therefore, it is necessary to consider the lagging impact of the economy on armed conflicts. Recent advances in deep learning, especially recurrent neural network (RNN), provide useful insights on how to address this problem. Theoretically, the RNN can solve long‐term dependency problems as its input contains the return value of the hidden layer, which enables the RNN to use historical data. However, Bengio, Simard, and Frasconi (1994) found that the RNN cannot successfully learn this knowledge. The long short‐term memory (LSTM), which is a special RNN structure, has been suggest to solve the problems of RNN gradient disappearance or explosion (Gers, Schmidhuber, & Cummins, 2000; Graves, Jaitly, & Mohamed, 2013; Shi et al., 2015). It has been proved very successful for numerous problems due to its capability of distinguishing between recent examples and early examples by assigning different weights for each while disregarding memory that it considers irrelevant to predict the next output. The LSTM is capable of handling long sequences of input compared with other RNNs that can only memorize short sequences (Nelson, Pereira, & Oliveira, 2017). Therefore, it can automatically reserve historical sequence information in its model structure for armed conflict prediction.

This neural network is based on the dynamic neural network theory, which can fully reflect the dynamic relationship between armed conflict events and the economy. The LSTM structure is shown in Fig. 2. It is characterized by the addition of valve nodes to RNN structure in each layer, which are forget gate, input gate, and output gate. These gates can be open or closed, to judge whether or not the output value in the memory status of model network has reached the threshold and add it to the compute process of current layer. The theory of LSTM is described as follows (Nelson et al., 2017; Shi et al., 2015):

Fig. 2

Technical flowchart of this study. OD in the figure represents observed data, and PD represents predicted data. In LSTM structure diagram, X_t is the input, and h_t is the hidden layer output. A refers to a module of neural network (Christopher, 2015).

Forget gate, which decides what information to abandon from cell state. The forget gate read the output from last layer and input from current state, and output a value between 0 and 1 by sigmod() function that is represented by σ. The output will be assigned to the current cell state , 1 indicates “preserved completely,” while 0 indicates “abandoned completely.” The calculation process of this layer is as follows:

Input gate, which decides how much new information to add to the cell state. It consists of two parts, one is the sigmoid() function layer, deciding the input information. Another is tanh layer, generating a new candidate vector, which would be added to state. The compute process is shown as following:

Next step is updating the old cell state from to . The compute process of this step is shown as following: where W is the weight matrices and the b is the bias vectors.

Output gate, which decides what values to export. This layer is based on the cell state, start with a sigmoid layer to determine which part of cell state to export, then deal with cell state using tanh (multiplied with the output of sigmoid layer), and finally, decide the output. The compute process of this layer is shown as following:

According to the analysis of LSTM steps above, LSTM effectively solves the problem of long‐term dependence of time series and can be used to simulate the relationship between armed conflict and the economy.

Prediction of Armed Conflict Based on LSTM

To achieve the goal of this study, the LSTM model was used to simulate the relationship between armed conflict and the economy. A technical flowchart was made to achieve the objectives of this study, as shown in Fig. 2.

Fig. 2 indicates that the following three steps should be completed:

Step 1: The data set should be built, including armed conflict events and feature dimensions that are used as input data for the model.

Step 2: LSTM, a machine‐learning algorithm on time series, was selected as the prediction method that should be employed.

Step 3: Different strategies should be applied to simulate the relationship between armed conflict and the economy. Two prediction strategies were employed in this study: a multiyear prediction and a yearly prediction.

In this article, Python (https://www.python.org/) was used to achieve the LSTM. The LSTM networks for armed conflict prediction were built using Keras (https://keras.io/), a python deep learning library. Some other python packages were also used to process the data and plotting, such as pandas (https://pandas.pydata.org/), numpy (https://numpy.org/), and matplotlib (https://matplotlib.org/). Besides, time step is important for the LSTM model. The choice of time step should reflect the temporal relevance of the data. With the increase of the time step, the number of layers of the LSTM spreading along the time dimension is increased, and the complexity of the network structure and the computational overhead are also increased. Through continuous trials, compared with other time steps, when the time step is 3, the training accuracy is the best. Therefore, the time step is 3 in this article. To overcome the shortcomings of other methods, pandas shift() function and the series to supervised() function were used to introduce the economic factors in the prediction, which was used to simulate the relationship between the economy and armed conflict. Besides, the sigmoid() function and tanh() function were used to determine which cell states are forgotten and which are added. In this way, the problem of long‐term dependency of RNN can be solved.

Data and Study Area

India was selected as the research area in this study. India is one of the riskiest regions for armed conflict. From 1989 to 2016, 14,464 armed conflict events occurred in India, which accounts for 11% of the armed conflicts in the world. Fig. 3 shows the temporal and spatial distribution of armed conflicts in India. Indian armed conflicts exhibit spatial differences and time fluctuations. Therefore, India is an ideal research area.

Fig. 3

Temporal and spatial distribution of armed conflicts in India.

The data set in this study includes armed conflict events data and economic indicators data. The data for armed conflict events were retrieved from the UCDP Georeferenced Event Dataset (UCPD GED), which is a leading conflict events data set (Eck, 2012). This data set is extensively employed in conflict research and is the UCDP's most disaggregated data set; it encompasses individual events of organized violence (phenomena of lethal violence that occurs at a given time and place). These events are sufficiently fine‐grained to be geo‐coded to the level of individual villages, with temporal durations disaggregated to single, individual days (Croicu, 2017; Sundberg & Melander, 2013). The UCDP GED, which covered the period 1989–2016—the maximum time period for which consistent data are available—was employed in this study.

The existing literature provides various economic variables that influence armed conflict events. An adequate set of economic indicators was built in this study based on existing literature to fully reflect socioeconomic conditions, including 4 major categories and 24 subcategories of economic indicators, namely, domestic macroeconomic indicators, international economic indicators, national income inequality indicator, and state fragility indicators according to the existing literature and expert recommendations. Annual average precipitation, annual average temperature, and DEM were selected as the background index to ensure the stability of the model. These three indicators maintain the same value in the time series to avoid disrupting the economy's role in conflicts. Detailed information about the 27 subcategories is listed in Table I.

Table I

Economic Indicators in This Study

	Economic Indicators
Indicators	Code	Code Meaning	Data Sources
Domestic macroeconomic indicators (Bukhari & Masih, 2016; Shahbaz, 2013)	EI 11	General government final consumption expenditure (current US$)	World Bank
	EI 12	Military expenditure (current LCU)
	EI 13	Total reserves (% of total external debt)
	EI 14	GDP per capita growth (annual %)
	EI 15	GDP per capita (current US$)
	EI 16	GDP growth (annual %)
	EI 17	GDP (current US$)
	EI 18	Claims on central government (annual growth as % of broad money)
International economic indicators (Lu & Thie, 2010; Sudula, 2012)	EI 21	Export value index (2000 = 100)	World Bank
	EI 22	Merchandise exports (current US$)
	EI 23	Export volume index (2000 = 100)
	EI 24	Import value index (2000 = 100)
	EI 25	Merchandise imports (current US$)
	EI 26	Import volume index (2000 = 100)
National income inequality indicator (Krieger & Meierrieks, 2015; Ted, 1968)	EI 31	Gini coefficient	World Income Inequality Database (WIID)
State fragility indicators (Kasasbeh, 2015)	EI 41	Security effectiveness	Center for System Peace
	EI 42	Security legitimacy
	EI 43	Political effectiveness
	EI 44	Political legitimacy
	EI 45	Economic effectiveness
	EI 46	Economic legitimacy
	EI 47	Social effectiveness
	EI 48	Social legitimacy
	EI 49	Human development index	United Nations Development Programme (UNDP)
Background indicators	Pre	Annual average precipitation	WorldClim
	Tem	Annual average temperature	WorldClim
	Dem	Digital elevation model	SRTM

RESULTS AND DISCUSSION

Armed Conflict and the Economy in India

Different economic indicators have different dimensions. To intuitively reflect the relationship between economic indicators and the frequency of armed conflict, normalization of the value of economic indicators was conducted; all ranged from 0 to 1,000. The trend of armed conflict frequency and the economic indicators are shown in Fig. 4.

Fig. 4

Trend of armed conflict frequency and the economic indicators.

Fig. 4 shows the correlation between armed conflicts and different economic indicators (see Table I). Subplot 1 (from top to bottom) shows the correlation between armed conflicts and domestic macroeconomic indicators (EI 11–EI 18), subplot 2 is the international economic indicators (EI 21–EI 26), and national income inequality indicator (EI 31), subplot 3 is the state fragility indicators (EI41–EI 49). As shown in three subplots, there is no significant correlation between armed conflicts and those economic indicators, which indicate that their affection on armed conflicts is not simply in linearity but in nonlinearity with lag memory effects.

Armed conflict showed a trend of initially increasing and then decreasing in Fig. 4. The number of armed conflicts was relatively low from 1989 to 1996, continuously increased from 1997, and reached a peak of 1,000 by the year 2000. Although the frequency of armed conflicts has been declining since 2000, it remained high and the risks were relatively high between 2000 and 2010. The frequency began to decline in 2011 and remained stable in the following years.

The macroeconomic situation in India shows an upward trend. However, the trends of the economic indicators significantly varied with no uniform relationship that was observed between armed conflicts and various economic indicators. For example, claims on central government and armed conflict show a synchronous trend from a macro point of view. However, GDP and armed conflicts generally show the opposite trend. Therefore, the economy's influence on terrorist attacks is complex. The deep learning method can be employed to simulate the relationship between armed conflicts and the economy.

The input data for LSTM model include armed conflict data and economic indicators from 1989 to 2016. Among them, 70% of the data were used as training samples and 30% of the data were used as test samples. Before simulation, the input data should be rebuild, and the MinMaxScaler() function and input_shape() function were used to normalize processing and build multidimensional input based on time step, respectively.

Multiyear Prediction

In multiyear prediction, armed conflicts from 1989 to 2009 were applied to predict the trends of armed conflict in the next few years. The rationality and robustness of the model should be verified before model prediction. The loss function was adopted to measure the inconsistency between the predicted value and the observed value of the model. This function can also be used to determine the model fit and whether under‐ or overfitting exists. One‐hundred simulations were performed to avoid the randomness of the results. The model evaluation results are shown in Fig. 5.

Fig. 5

Model train and test loss. These two curves eventually converged on the same level and remained stable, which indicates that the results of the simulation do not appear overfitting and underfitting.

Fig. 5 shows that the results of the simulation do not appear overfitting and underfitting. Therefore, LSTM was used to predict the frequency of armed conflict from 2010 to 2016. The results are shown in Fig. 6.

Fig. 6

Simulation results of armed conflict in multiyear predictions.

Fig. 6 indicates that the LSTM model provides a suitable prediction for the trend of armed conflicts on the whole. As shown in the figure, the predicted values between 1995 and 1999 correspond well with the actual situation, this is also the case with years after 2001. However, the predicted values of the armed conflict frequency are biased in some years. For instances, in 2000, the predicted values of 100 times simulation are much lower than the observed value. To quantitatively analyze the prediction accuracy, we compare the accurate data with the prediction data and calculate the prediction accuracy of each year using the following formula:

We extracted the results of each simulation and calculated the simulation accuracy of each year by formula (7). The results are shown in Table II.

Table II shows that the accuracy of the model prediction significantly varies from year to year and ranges from 83.1% to 98.8%. The average simulation accuracy is 90.5%. The Root Mean Square Error (RMSE) of each simulation was calculated and used to measure the deviation between the predicted data and the observed data. The average RMSE is 61, which is large. In general, LSTM can accurately predict the future trend of armed conflicts in the future. However, some bias exists in forecasting the frequency of occurrence. In general, there are three ways to correct bias: increasing the sample size, adjusting the simulation strategy, and modifying the model parameters. In this study, the adjustment strategy was used in this article to correct the bias that is yearly prediction.

Table II

Prediction Results and Accuracy from 2010 to 2016 in a Multiyear Prediction

	Frequency of Armed Conflict
Year	Observed Data	Average Predicted Data	Prediction Accuracy
2010	784	672	85.7%
2011	403	335	83.1%
2012	425	420	98.8%
2013	344	386	87.8%
2014	394	416	94.4%
2015	323	293	90.7%
2016	417	447	92.8%
Average	–	–	90.5%

Yearly Prediction

To solve the problem of large deviations in forecasted results, the yearly prediction was used to predict the number of armed conflicts from 2000 to 2016. First, the frequency of armed conflict in the period 1989–2009 was used to simulate the relationship between the economy and armed conflict, which was employed to predict the armed conflict number in 2010. Second, the observed data of 2010 were added to the model, and the observed data from 1989 to 2010 were used to predict the frequency of armed conflicts in 2011. This procedure was performed for each year. Last, the number of armed conflicts in 2016 was predicted with the frequency of armed conflict in the period 1989–2015. To avoid the randomness of the results, the model was simulated 100 times. The simulation results for each year are presented in Fig. 7.

Fig. 7

Simulation results of armed conflict in each year.

Fig. 7 shows that the predicted results are consistent with the real situation and LSTM adequately predicted armed conflicts each year since 2010. The prediction results are shown in Table III.

Table III

Prediction Results and Simulation Accuracy from 2010 to 2016 in a Yearly Prediction

	Frequency of Armed Conflicts
Year	Observed Data	Average Predicted Data	Average RMSE	Prediction Accuracy
2010	784	668	115	85.2%
2011	403	483	80	80.1%
2012	425	420	29	98.8%
2013	344	359	25	95.6%
2014	394	376	23	95.4%
2015	323	304	20	94.1%
2016	417	405	14	97.1%
Average	–	–	44	92.4%

Formula (3) was used to calculate the simulation accuracy, whose values range from 80.1% to 98.8%. The average accuracy of the yearly prediction is 92.4%. The accuracy of the yearly prediction is higher than that of the multiyear forecast. The average RMSE for the yearly prediction is 44, which is less than the average RMSE of the multiyear forecast. The average RMSE for each year shows an annual decreasing trend, which indicates that the prediction results are similar to the observed data.

Table III also shows that the variation trend of the average RMSE is inconsistent with the prediction accuracy for each year. For the yearly prediction, the average RMSE shows an annual decreasing trend but the prediction accuracy is random. For example, the prediction accuracy for 2012 is 98.8%, which is the highest prediction compared with other years. However, the average RMSE for 2012 is 29, which is higher than the predictions of many other years. The average RMSE for 2016 is 14, which is the smallest prediction of all years. However, the prediction accuracy in 2016 is 97.1%, which is lower than that in 2012. The main reason for the inconsistency between the trend of average RMSE and prediction accuracy is that the predicted data may be higher or lower than the observed data for each simulation, and the average predicted data can cancel some errors. The RMSE represents the deviation between each predicted data and the observed data and is positive. Therefore, the RMSE of the year with the high prediction accuracy is likely to be large.

Discussion

LSTM is a time‐RNN, which is suitable for predicting events with relatively long intervals and delays in time series. Compared with other methods, the LSTM can solve the problems of long‐time dependence and multielement input. In this study, LSTM was used to simulate the relationship between the economy and armed conflict. The results show that the average simulation accuracy is 90.5% and 92.4%, respectively, under multiyear forecast and yearly forecast. Although LSTM can well simulate the relationship between armed conflict and economy, there are still some shortcomings. For example, LSTM cannot quantitatively calculate the contribution rate of each factor. Therefore, it is unknown which economic factor has the greater impact on the occurrence of armed conflict. Besides, LSTM also has the problems of computationally time‐consuming and unable to process a large number of sequences, which need to be improved in the future.

Although the number of armed conflicts was predicted, it is more scientific to treat the results as a level of risk from an application perspective. The frequency of armed conflicts is not enough for the country to prevent the occurrence of armed conflicts, because it is not known when and where the armed conflicts may occur. We can treat this risk as a degree of harm. The more frequently armed conflicts occur, the greater the harm they generally cause. For example, in some years when armed conflicts frequently occur, conflict risk levels are high, and prevention should be strengthened compared with other years.

Two strategies were applied in this study: a multiyear prediction and a yearly prediction. The results indicate that the yearly prediction is more accurate than the multiyear prediction. The LSTM structure in Fig. 2 indicates that the simulated data will be added to predict subsequent simulated data in multiyear prediction. The observed data were added in the next simulation in yearly prediction. Therefore, the yearly prediction is more accurate. Longer time series data can train the model more accurately, and the result is more accurate. Theoretically, the effect will be better if the data are available from the month‐scale simulation. Besides, adequate and effective economic indicators will also improve simulation accuracy. Therefore, building a more adequate and effective economic indicator system is the next step that needs to be improved. The method is feasible in general. If future economic data can be predicted, the future state and trend of armed conflict can be predicted with LSTM and future economic data, which is very important for the Indian government.

CONCLUSIONS

A relationship between armed conflict and the economy exists. In this study, LSTM, which is a deep learning method, was used to simulate the relationship between armed conflict and economy. To fully reflect the economic situation of India, 4 major categories and 24 subcategories of economic indicators were chosen. The degree of conformity of the simulation results with true values is used to prove that the economy is the main factor in the occurrence of armed conflicts. The prediction accuracy confirmed that the risk of armed conflicts can be predicted by economic indicators.

Two simulation strategies were employed in the study: multiyear prediction and yearly prediction. In the multiyear prediction, the frequency of armed conflicts from 1989 to 2009 served as the training data to predict the number of armed conflicts from 2010 to 2016. In the yearly prediction, the frequency of armed conflicts in the period 1989–2009 was used to predict the number of armed conflicts in 2010. Data from 2010 to 2015 were gradually added to the model to predict the number of armed conflicts in the next year. The number of armed conflicts in 2016 was predicted with the frequency of armed conflict in the period 1989–2015. To avoid the randomness of the results, the two strategies were simulated 100 times. The results show that both strategies can adequately simulate the relationship between the economy and armed conflicts. The average simulation accuracy is 90.5%, and the average RMSE in the multiyear prediction is 61. In the yearly prediction, the average simulation accuracy is 92.4%, and the average RMSE is 44. Therefore, the accuracy of the yearly prediction is higher than that of the multiyear forecast. The average RMSE for each year shows an annual decreasing trend, which indicates that the prediction results are similar to the observed data.