INTRODUCTION Fruits are significant for a healthy diet since they are rich in vitamins, minerals, and fibre. It improves farmers' socioeconomic conditions and has also become a source of better income for many underprivileged groups.Citrus fruits are one of the healthiest fruits which are nutritional powerhouses. It is mostly produced in Spain, Brazil, China, the United States, Mexico, and India. Citrus fruit’s export from India is a booming enterprise, owing to the country's favourable climate and soil conditions. The mandarin (Citrus reticulata), sweet orange (Citrus sinensis), and acid lime (Citrus aurantifolia) are the most important commercial citrus species in India, accounting for 41, 23 and 23 percent of all citrus fruits produced in the country respectively. After mango and banana, India's citrus business is the country's third largest fruit industry. India is the ninth largest orange producer in the world, accounting for 3% of worldwide orange production. Gujarat is holding third rank in terms of citrus production (National Horticulture Board, 2020-21). Many academics have been utilising statistical models to forecast area, production, and productivity of various crops in recent years. Dhaikar and Rode (2014) used artificial neural network approach for predicting agricultural crop yield. Guo and Xue (2014) applied artificial neural network for forecasting crop yield and made comparison between spatial and temporal models. Hamjah (2014) used the Box- Jenkins ARIMA model to anticipate significant fruit crop production in Bangladesh. Time series analysis for pineapple production in Bangladesh was computed by Hossain and Abdulla (2015). Hossain et al. (2016) used various statistical methods to anticipate banana production in Bangladesh. Dasyam et al. (2016) applied statistical modelling to predict area, production and productivity of potato in West Bengal. Kumari et al. (2016) used various statistical models to forecast pigeon pea yield in the Vanaransi region. Kumari et al. (2017) looked at forecasting models for predicting pigeon pea pod damage in the Varanasi region. Using various statistical models, Rathod and Mishra (2018) estimated mango and banana yield in Karnataka. Aguilar et al. (2020) used artificial neural network model for predicting greenhouse tomato yield and aerial dry matter. In Gujarat, Kumar and Kumari (2021) predicted the area, production, and productivity of sapota. Computed a trend study of minor millet area, production, and productivity in India. Unjia et al. (2021) looked into the trend of maize area, production, and productivity in India. Yield forecasting is a way to support policy decisions for fruit crops in order to introduce confidence amongst the farming community for their socio-economic issues. Thus, modeling and forecasting the area, production and yield of citrus fruit crop over the years are of much practical importance. The present study was taken up in order to compare the performance of different statistical models i.e., Autoregressive Moving Average Models (ARIMA) and Exponential smoothing for predicting area, production and yield of citrus in Gujarat. Traditional forecasting methods in India continue to cause problems in predicting crop area, production, and productivity, as well as market price of agricultural commodities. This research has a broad scope in that it aims to develop adequate statistical models for predicting data on area, production, and productivity of various crops, with the proper model assisting in the execution of some forecasting policies. METHODOLOGY Time series secondary data on area, production and yieldof Citrusin Gujaratwere collected from Directorate of Horticulture, Govt. of Gujarat from 1991-92 to 2016-17. Analytical framework: In the present study, time series forecasting models i.e., Exponential smoothing (ES) andAutoregressive integrated moving average (ARIMA), were used to compare their ability for predicting future behavior of area, production and yield ofcitrus in Gujarat. Analysis was done by RStudio version 3.5.2 software. Time series models used in the present investigation were: Exponential Smoothing (ES) model: Smoothing techniques are used to reduce irregularities (random fluctuations) in time series data. One of the most successful univariate time series forecasting technique is the exponential smoothing (ES) to produce a smoothed time series. In this technique, forecasts are weighted averages of past observations, with the weights decaying exponentially as the observations get older. In other words, recent observations are given relatively more weight in forecasting than the older observations. Exponential smoothing method is classified according to the type of component (trend and seasonality) presented in the time series data. In the present study, based on time series data, only two exponential smoothing methods are used i.e., simple exponential and double exponential smoothing technique. Simple Exponential Smoothing (SES): This method is suitable for forecasting data with no trend or seasonal pattern, although the mean of the data may be changing slowly over time. Forecasts are calculated by taking weighted averages of most recent observation and most recent forecast, where the weights decrease exponentially as observations come from further in the past. Forecast equation y ̂t+1/t =〖 l〗_t Level equation l_t=αy_t+(1-α)l_(t-1) Simple exponential smoothing has a flat forecast function, and therefore for longer forecast horizons, y ̂t+h/t =( y) ̂t+1/t = l_t Holt's linear trend(double) exponential smoothing method: Holt (1957) extended simple exponential smoothing to allow forecasting for those data which exhibit trend. This method involves a forecast equation and two smoothing equations (one for the level and one for the trend): Forecast equation y ̂t+1/t=l_t+hb_t Level equation l_t=αy_t+(1-α)(l_(t-1)+b_(t-1)) Trend equation b_t=〖β(〖l_t-l〗_(t-1))〗_ +(1-β) b_(t-1) Where y_t, y ̂t are observed and predicted value of series at time t l_(t )and〖 b〗_t are estimate of the level and trend (slope) of the series at time t α,β are the smoothing parameter for the level and trend, 0≤α,β ≤1 Initialisation: The application of every exponential smoothing method requires the initialisation of the smoothing process. For simple exponential smoothing we need to specify an initial value for the level, l_0. Similarly double exponential smoothing involves initial value trend component〖 b〗_0 also. In exponential smoothing, the method for obtaining the optimal values of smoothing parameters α and β is an iterative process which is chosen either by trial-and-error method or by some software like MINITAB, E Views, SPSS etc. which use an algorithm to select the value of the weights that minimizes mean square error for in-sample forecasts. Autoregressive Integrated Moving Average (ARIMA) model: ARIMA (Box and Jenkins 1970) models provide another approach to time series forecasting. Exponential smoothing and ARIMA models are the two most widely-used approaches to time series forecasting, and provide complementary approaches to the problem. While exponential smoothing models were based on a description of trend and seasonality in the data, ARIMA models aim to describe the autocorrelations in the data. ARIMA is one of the most traditional methods of non-stationary time series analysis. Usually time series, showing trend or seasonal patterns are non-stationary in nature. In such cases, differencing and power transformations are often used to remove the trend and to make the series stationary. Box-Jenkins ARIMA, has been successfully applied in many time series forecasting and is a good tool to develop empirical model which is linear combination of its own past values, past errors (also called shocks or innovations). ARIMA model allows Yt to be explained by its past, or lagged values and stochastic error terms. The non-seasonal ARIMA (p, d, q) model can be written as: If then Where P: order of the autoregressive part; d: degree of differencing involved; q: order of the moving average part. & t: Differenced data series and white noise & : Autoregressive and moving average coefficient The main stages in setting up a Box-Jenkins forecasting model are model identification, estimating the parameters, diagnostic checking of residual and forecasting (Box and Jenkins 1970). RESULTS In this study, area, production and yield (yield) of Citrus crop were analyzed by Exponential smoothing model and ARIMA model. The empirical findings of citrus crop are as follow: Forecasting of area, production and yield for Citrus Forecasting of area for citrus. Fig. 1 illustrate chart series of area dataset for citrus from1991-92 to 2016-17. Also, the characteristics (basic statistics) of the data set used were presented in the Table 1. Exponential smoothing (ES) model: In case of fitting exponential smoothing model, the performance of ETS(A, A,N) holt linear method was found to be the best out of all. The results were shown in Table 2. Table 2 shows that the estimate of alpha was found to be significant only. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.24. The forecasted value of Citrus area in Gujarat for the year 2017-18 was obtained as 46.50(‘000’ Hectares) with confidence interval 43.05to 49.96. Autoregressive Integrated Moving Average (ARIMA) Model: In case of fitting ARIMA model, out of various ARIMA models with different value of p, d and q, the performance of ARIMA (1, 1, 0) with drift was found to be the best. The results were given in Table 3. Table 3 reveals that the estimates of all parameters were found to be statistically significant. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.92. The forecasted value of citrus area in Gujarat for the year 2017-18 by ARIMA (1, 1, 0) with drift was obtained as 46.39 (‘000’ Hectares) with confidence interval 43.06to 49.72. Forecasting of production for Citrus: Fig. 2 illustrate chart series of production dataset for citrus from 1991-92 to 2016-17. Also, the characteristics (basic statistics) of the data set used were presented in the Table 4. Exponential smoothing (ES) model: In case of fitting exponential smoothing model, the performance of ETS(A,A,N) holt exponential method was found to be the best out of all. The results were shown in Table 5. Table 5 shows that the estimate of alpha was found to be significant. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.48. The forecasted value of Citrus production in Gujarat for the year 2017-18 was obtained as 606.37 ('000' MT) with confidence interval 531.98 to 680.76. Autoregressive Integrated Moving Average (ARIMA) Model: In case of fitting ARIMA model, out of various ARIMA models with different value of p, d and q, the performance of ARIMA (0,1,0) with drift was found to be the best. The results were given in Table 6. Table 6 reveals that the estimate of alpha was found to be statistically significant. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.42. The forecasted value of citrus production in Gujarat for the year 2017-18 by ARIMA (0, 1, 0) with drift was obtained as 606.55('000' MT) with confidence interval 536.25 to 676.84. Forecasting of yield for citrus: Fig. 3 illustrate chart series of yield dataset for citrus from 1991-92 to 2016-17. Also, the characteristics (basic statistics) of the data set used were presented in the Table 7. Exponential smoothing (ES) model: In case of fitting exponential smoothing model, the performance of ETS(M,N,N) simple exponential smoothing method was found to be the best out of all. The results were shown in Table 8. Table 8 shows that the estimate of alpha was found to be significant. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.37. The forecasted value of citrus yield in Gujarat for the year 2017-18 was obtained as 13.04 ('000' MT) with confidence interval 8.00 to 18.09. Autoregressive Integrated Moving Average (ARIMA) Model: In case of fitting ARIMA model, out of various ARIMA models with different value of p, d and q, the performance of ARIMA (1, 0, 0) with constant was found to be the best. The results were given in Table 9. Table 9 reveals that the estimate of both parameters was found to be statistically significant. Also, residual autocorrelation was non-significant as per Box-Ljung test statistics probability value 0.31. The forecasted value of citrus yield in Gujarat for the year 2017-18 by ARIMA (1, 0, 0) with constant was obtained as 12.01(MT/Ha.) with confidence interval 6.49 to 17.52. Table 10 shows the performance of different models for predicting area, production and yield of citrus. Area, production and yield of citrus was best explained by ARIMA model with forecasted values 46.39 (‘000’ Ha.), 606.55 (‘000’ MT) & 12.01 (MT/Ha.) respectively.