Comparison of Forecasting Methods For Agriculture

Comparison of forecasting methods for agriculture
CHATPTER 1
1.
Introduction
The field of forecasting is concerned with making statements about matters that are currently unknown. The terms forecast, prediction, projection, and prognosis are interchangeable as commonly used. Forecasting is also concerned with the effective presentation and use of forecasts. Useful knowledge comes from empirical comparisons of alternatives and this entry is concerned primarily with evidence-based or scientific procedures. Before forecasting, one should consider whether it is necessary. Forecasting is needed only if there is uncertainty; a forecast that the tide will turn is of no value. Forecasts are also unnecessary when one can control events. Forecasting should not be confused with planning. Whereas planning is concerned with what the planner thinks the future should be like, forecasting is concerned with what it will be like. Managers should start by planning. Forecasting procedures are then used to predict outcomes for the plans. If the managers do not like the forecasts, the planning and forecasting processes can be repeated until a plan is found that leads to forecasts of acceptable outcomes. The best plan can then be implemented and actual outcomes monitored so that the feedback can be used in the next planning period. The Methodology Tree for Forecasting (Figure 1) is a classification schema of all forecasting methods organized on the basis of the source of the knowledge the forecaster has about the situation. Some methods use primarily judgmental or qualitative knowledge while others require statistical data. There is an increasing integration in the use of judgment and statistics in the procedures as one follows the Tree down.
The most common way to make forecasts is to ask experts to think about a situation and predict what will happen. If experts forecasts are derived in an unstructured way the approach is referred to as unaided judgment. It is fast, can be inexpensive when few forecasts are needed, and can be appropriate when small changes are expected. It is most likely to be useful when the forecaster knows the situation well, makes frequent forecasts, and gets good feedback about the accuracy of his forecasts, as is the case with short-term weather forecasting and sports betting.
Expert forecasting refers to combining forecasts obtained from experts using validated structured techniques. Which method is most appropriate depends on time constraints, dispersal of knowledge, access to experts, expert motivation, and need for confidentiality. To
IEM, MED, NIT Calicut Page 1
use expert forecasting methods, one should obtain the services of between five and twenty diverse experts who each have relevant information. Pre-test questions to elicit forecasts from the experts, and specify procedures for combining the forecasts obtained from the experts (e.g., use the median) in advance.
Fig. 1: Methodology Tree for Forecasting
Consider also the Delphi method for combining the forecasts of experts. Delphi involves at least two rounds of anonymous interaction between experts. An administrator summarizes individual forecasts and arguments and reports this feedback to participants after each round. In the light of the feedback, the participating experts provide revised forecasts and further reasoning. The Delphi forecast is a combination of the final round forecasts. Prediction markets can be useful for combining forecasts to provide continuously updated numerical or probability forecasts. In a prediction market, anonymous participants reveal information by trading contracts whose prices reflect the aggregated group opinion. Incentives to participate in a market may be monetary or non-monetary. Although prediction markets seem promising, to date there has been no published meta-analysis of the accuracy of prediction market forecasts.
IEM, MED, NIT Calicut
Page 2
The structured analogies method uses information about similar situations to obtain forecasts. Experts identify situations that are analogous to a target situation, describe similarities and differences to the target, and then derive an overall similarity rating. The outcome or decision implied by each experts top-rated analogy is the structured analogies forecast from that expert.
Judgmental bootstrapping is a method for deriving a forecasting model by regressing experts forecasts against the information the experts used to make their forecasts. The method is useful when expert judgments have predictive validity but data are scarce (e.g., forecasting new products) and outcomes are difficult to observe (e.g., predicting performance of executives). Once developed, judgmental bootstrapping models are a low-cost forecasting method. A meta-analysis found judgmental bootstrapping to be more accurate than unaided judgment in 8 of 11 comparisons. Two tests found no difference, and one found a small loss in accuracy. The typical error reduction was about 6%
Expert systems are forecasting rules derived from the reasoning experts use when they make forecasts. They can be developed using knowledge from diverse sources such as surveys, interviews of experts, or protocol analysis in which the expert explains what he is doing as he makes forecasts. A meta-analysis on the predictive validity of the method found that expert systems were more accurate than unaided judgment in six comparisons, similar in one, and less accurate in another. Expert systems were less accurate than judgmental bootstrapping in two comparisons and similar in two.
Role playing involves asking people to think and behave in ways that are consistent with a role and situation described to them. Role playing for the purpose of predicting the behaviour of people who are interacting with each other is called simulated interaction. The decisions made in the simulated interactions are used as forecasts of the actual decision.
Conjoint analysis is a method for eliciting peoples preferences for different possible offerings (e.g. for alternative mobile phone designs or for different political platforms) by exposing people to several combinations of features (e.g. weight, price, and screen size of a mobile phone.) The possibilities can be set up as experiments where variations in each
Page 3
variable are unrelated to variations in other variables. Regression-like analyses are then used in order to predict the combination of features that people will find most desirable.
Page 4
CHAPTER 2
2. Importance of agricultural forecasting.
Economic forecasting in agriculture has some features in common with business forecasting and with macroeconomic forecasting. But over time, it has developed a focus of its own. During the second quarter century, the profession shifted toward prediction, broadly defined, including use of econometric techniques for estimating elasticities and forecasting prices. The third quarter century, from 1975 onwards, has been characterized by research on policy, trade and the global economy and expansion to environmental and resource problems. Throughout the entire period, and more markedly of late, explanation of past behaviour has been the dominant focus of agricultural supply modelling, which is the area to which most agricultural forecasting belongs.
Because an assured food supply is important to national security, governments have attempted to quantify agricultural production and to exert some control over it. In the beginning, simply collecting and tabulating data on the current agricultural situation was a major challenge, and agricultural statisticians played a major role in the development of statistical methods. Data revision was frequent. Estimates of production, for example, were subject to revision after a new census had been tabulated. The large number of Situation reports or similarly titled publications indicates the fascination of agricultural statisticians with estimating the current status of a data series. The nature of agricultural production and the historical relations among the different groups of participants in agriculture make agriculture different from most economic activity. Most product is unbranded and sold in markets where individual suppliers have no say in price determination. Both nature and government policy can have a major impact on a farmers production and profits. Farmers and others connected with agriculture are used to receiving technical and economic information from publicly supported institutions.
2.1. Characteristics of agricultural production Agricultural production is unusual compared with most business activity in its strong dependence on biological processes. Farmers have minimal ability to alter the rate of development of a crop or animal. Second, for most commodities, the production cycle is measured in months or years. Other features impose dynamic structure, especially on prices:
Page 5
seasonal impacts on production, high cost of adjustment once production is underway and the need to carry inventory. Estimation of leading indicators therefore became a major part of short-term agricultural production forecasting, dominating any work on price forecasting. The estimation of leading indicators was a natural extension of the data gathering activity concerning current production or inventories. For example, estimation of acres planted to spring wheat is a good indication of harvested acreage. In no other sector has leading indicator analysis found such long-term and widespread use.
Agricultural production appears to meet the four conditions laid down for good forecasts by econometric methods there should be strong causal relationships, relations should be capable of being measured accurately, causal variables should change substantially and it should be possible to forecast changes in causal variables. Unfortunately, econometric methods do poorly at forecasting agricultural production and prices. The most likely reason is the great influence on production of random shocks. Relative to most manufacturing activity, agriculture is greatly influenced by unpredictable random events such as droughts, hoods and attacks by pests. The consequence of these shocks on production can be assessed reasonably well after they have occurred, which is useful in making post-harvest production estimates, but not pre-harvest forecasts.
2.2 Producers of agricultural forecasts The predominant forecaster of production, prices and trade of agricultural commodities and inputs in most countries is central government. The Economic Research Service of the United States Department of Agriculture (USDA-ERS) contains the largest agglomeration of agricultural economists and produces the greatest number of agricultural forecasts. Government commodity specialists are the main providers of outlook information in Australia, Canada and the US. Reports on the situation and outlook for commodity and input markets at local, national and world levels are issued from one to twelve times a year depending on commodity and country. Some agencies issue regular medium-term forecasts (2-5 years ahead). For example, Agriculture Canada has issued Long-term projections are generally issued only irregularly, and usually for groups of commodities. Although governments publish many forecasts, often as regular series, they also make many forecasts solely for internal use, for example, the USDA forecasts of the budgetary cost of the farm program medium-term outlook reports twice a year since.
Page 6
Other public agencies, from the Food and Agricultural Organization of the United Nations to regional or provincial governments, also produce forecasts. University faculty and (in the US) extension economists prepare forecasts for general release as part of short-term outlook programs for local farmers and agribusinesses. They may also present forecasts in scholarly publications; these usually have a methodological focus.
Private companies that process or trade commodities or supply inputs produce forecasts for in-house use, typically with relatively simple models combined with judgment. They are probably closest to business forecasters in both approach and objectives. Private consultants also produce forecasts for sale, most frequently as adjuncts to large-scale macroeconomic models. Farmers practically never produce formal forecasts, though most of them doubtlessly form a judgment about future outcomes of their business choices.
2.3 Users of agricultural forecasts Farmers may rarely make forecasts, but they form the largest group of users. They need to make production and marketing decisions that may have financial repercussions many months in the future. Short-run commodity outlook forecasts, at least in the US, have tended to emphasize production and inventory information. Farmers have more use for price forecasts. Once committed to a product, farmers are price takers. They produce goods that are homogeneous or highly substitutable with the goods of their competitors, who may either be their neighbours or live halfway round the world. They have no concern with problems common in manufacturing, such as the amount of sales of a branded product or what quantity of a specific model to keep in inventory. But farmers, especially those in developed countries, must also be concerned with the ways in which changes in government policy will alter their business conditions. Agricultural journalists represent a second kind of audience for commodity forecasts. They are not users in the sense of being makers of decisions based on forecast information. They provide an indirect way for readers and listeners (mainly farmers) to receive outlook forecasts. Processors of food and fibre, and others in the marketing chain, need forecasts to aid in their purchasing and storing decisions. They too would probably like price forecasts, but would be able to make greater use of production forecasts in their decisions than would farmers. Larger businesses also supplement public forecasts with their own in-house ones.
Page 7
Governments in many countries intervene in agricultural production to protect domestic agriculture and provide food security. For this they need two kinds of information. First, for legislation and, to a much lesser extent, for program implementation, governments need to know the consequences of different policy choices on different groups in society. Agricultural economists have been especially willing, over the last 30 years, to build ever larger models to provide answers to policy questions. Emphasis has been placed on comparing proposed policies via simulations, which has measurably assisted legislators. Forecasts of output and prices are conditional on the policy actually selected. To date, efforts to forecast which policy will be selected have been minimal.
Neither have government or academic economists done much to evaluate a models ability to forecast the actual consequences of an adopted policy. Second, in monitoring the progress of farm programs designed to control supplies or support prices, governments would like to know about the effectiveness of the program and anticipated budget outlays.
Page 8
CHAPTER 3
3. Data collection for analysis
Agricultural data is not very difficult to obtain. But, getting the authentic data was very important for our analysis. Data from different sources were collected, sorted, analysed and finally we had narrowed down our search to the data published by the central government of India, Ministry of Agriculture. The data we collected consisted of the past history of the produce of about 25 crops over a span of 25 years from 1984 to 2008. After further analysing the data we selected 10 crops for the further analysis. The selection of the data was purely based on the fact that some of the crops did not have a consistent set of data ie. for some crops the data over the years were not present.
The data analysis part consisted of two main parts. First, it was to study the property of the data. For the given model the data was assumed to follow the first order auto-correlation. The stage consisted of determining the parameters required for the forecasting models and to forecast one period into future and compare the models. For the comparison we have selected three models. Firstly, Exponentially Weighted Moving Average, secondly, Trend Adjusted moving Average and finally, the Mean Square Optimal method.
The parameters required for each was determined using either analytical of by using simulation. The data collected is being tabulated in the table below.
Page 9
Table 1: Agricultural data over past 25 years (in mega tonnes)

crop year 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 rice
17427944 19081147 18314686 17013045 21171447 21969341 22597604 22721384 22074476 24449500 24917312 23404729 24877437 25106636 26218740 27361764 25890853 28546322 21808324 27050186 25332886 28042829 28341887 29476025 30246312
sugarcane
3380547 3307590 3334036 3613848 3820620 3942964 4380542 4681091 4932660 4428327 4460179 5350972 5458947 5390201 5428681 6026647 5781769 5716715 5740895 5551122 4517297 4604244 5460342 6904174 6725632
wheat
6447852 6248398 6671234 6284262 6546232 7672064 7305245 8087016 8195579 8424183 8820980 9745848 9165961 10271064 9795029 10557335 11367806 10327497 10820910 9719472 10708479 10164358 10251551 11245019 11671546
milk
11850637 12587400 18314686 12898588 13206280 14195790 14513971 14323162 14785200 15407577 16247236 17917065 18751729 19334646 20186293 21112866 21692286 22678299 23232744 23965510 25063911 26008965 27163809 27928543 30419550
vegetables
4128300 4203360 4447305 4503600 4334715 4184595 4240890 4334715 3865590 4147065 4053240 3788653 3753000 3753000 4240890 5072179 5372419 6631551 5100327 6518961 3597757 4151512 5156772 5469246 5892585
potato
1565582 1627722 1318852 1642566 1820911 1936852 1925393 1942138 2351405 2437367 2274581 2266681 2443399 3231954 2263357 2951660 3345146 2933998 3237019 3014991 3696970 3811475 3827306 3731788 4602900
millet
1601937 1203988 1360891 1117846 1828702 1733080 1681365 1317172 1974488 1374719 1657749 1384728 1769808 1676829 1656078 1391515 1626972 1832010 1020485 2394778 1756474 1703087 1676096 2064876 1841508
maize
894614 704062 804661 606250 872046 1022827 899817 721416 848155 713468 611955 655484 787138 783885 821034 840428 873731 979589 771997 1181440 1073339 1120272 1139224 1550386 1442042
groundnuts coconut
2869123 2238634 2607717 2540184 4372493 3633177 3338196 3152164 3857443 3518262 3641368 3405527 3936811 3311295 4107969 2840428 2920525 3195190 3771997 3702426 3045803 3672741 4113922 4226948 4144204 464517 454913 428504 488556 573932 628919 653881 678300 755264 804373 893719 870274 877268 813055 791350 780497 755174 784114 806724 780497 757887 798494 921583 985253 985253
Page 10
CHAPTER 4
4. Study of properties of data
Suppose we have a time series regression model relating a "dependent" time series {yt } to the independent" time series {x1t } , . . . , {xpt }. The model is
where {t } is a time series of "errors", or "disturbances". Such models are useful for both explanatory and forecasting purposes. The parameters 0 , 1 , . . . , t may be estimated by least-squares. In practice, it often happens that the errors are not independent (as assumed in standard regression models) but instead are autocorrelated. Such error autocorrelation, or "serial correlation", has many undesirable but correctable consequences (e.g., the leastsquares estimates sub-optimal, standard confidence intervals for are incorrect, the error term is forecastable). Thus, it is highly desirable to try to detect error autocorrelations.
The Durbin-Watson Test for serial correlation assumes that the t are stationary and normally distributed with mean zero. It tests the null hypothesis H0 that the errors are uncorrelated against the alternative hypothesis H1 that the errors are AR (1). Thus, if s are the error autocorrelations, then we H0 =0 (s > 0), and H1 = s=s for some nonzero with || < 1. To test H0 against H1 , get the least squares estimates statistic is and residuals e1 , . . . , en . The test
Note that ignoring "end effects", we have d = 2(1 - r 1), where r 1 is the sample ACF of the residuals at lag 1. If the errors are white noise, d will be close to 2. If the errors are strongly auto correlated, d will be far from 2. The exact procedure for deciding whether a given value of d is significant is somewhat complicated. In some cases, the test can be "inconclusive," i.e., H is neither accepted nor rejected.
Since its development in 1951, the test has been found to be extremely useful, especially for the analysis of economic time series. It does, however, suffer from a number of shortcomings, some of which are as follows. First, the form of the model (i.e., the dimension p and the
explanatory variables x1t , . . . , xpt ) is assumed known. In practice, this is rarely the case, and instead a data based procedure must be employed to "identify" the model. Second, the test is sometimes inconclusive, as mentioned above. Third, the AR (1) alternative hypothesis is by no means the only way in which the null hypothesis may fail. Suppose, for example, the errors are in fact MA(1), or perhaps even some nonstationary series such as a random walk. The Durbin-Watson test can have very low power against such alternatives (i.e., it can fail to detect them).
There are a number of potential stochastic processes that can be assumed for the yield process, ranging from a simple independent and identically distributed (IID) process to a nonstationary process. In reality, there is empirical evidence that few real-world processes exhibit pure IID behaviour (Alwan, 2000). One correlative yield process that has been frequently studied in the supply-chain literature is the first-order autoregressive model, AR(1) (Chen, Drezner, Ryan, & Simchi-Levi, 2000a, 2000b; Kahn, 1987; Lee, Padmanabhan, & Whang, 1997, 2000). The AR(1) process is often studied because pure first-order correlative effects commonly occur in real-world processes (Alwan, 2000). Furthermore, even for processes that do not follow a pure AR(1) process, an AR(1) model is a flexible model that can reasonably capture lagged effects so commonly found in demand processes. In the following subsections, we describe the AR(1) process in more detail, present the forecasting models to be considered.
4.1 Demand process Suppose the system is faced with an AR(1) demand process given as follows: (1) where dt is the observed demand in period t, > 0, | | < 1, and {t} is an independent and identically distributed normal (IIDN) process with mean 0 and variance 2e. The condition of | | < 1 ensures that the process is stationary. It is useful to note that = and = (1 ) 1
Page 12
In practice, if
is unknown, we may first obtain unbiased estimates of demand parameter, ,
and then substitute the demand parameter with the corresponding parameter estimate.
As noted earlier, the AR(1) model is widely used in supply-chain management research (e.g., Chen, Ryan, & Simchi-Levi, 2000b; Lee, So, & Tang, 2000). One of the primary reasons for the focus on the AR(1) model is that it possesses good dynamic characteristics. Namely, by varying the parameter, we are able to study the effects of processes which are random, nonrandom but stationary or even nonstationary processes. This flexibility allows us to gain practical insights for many real demand patterns.
By varying the values of , one can represent a wide variety of process behaviours. When = 0, we have an IIDN process with mean s and variance . For _1 < < 0, the process is < 1,
negatively correlated and will exhibit period-to-period oscillatory behaviour. For 0 <
the demand process will be positively correlated which is reflected by as wandering or meandering sequence of observations. As approaches | |1, the process approaches
nonstationary behaviour, most notably, the random walk model ARIMA(0,1, 0) is a special case of the AR(1) model when = 1. As pointed out by Graves and Willems (2000),
varying a stationary demand model is an important exercise for gaining fundamental insights into the relationship between variables such as inventory and orders relative to demand characterization.
Forecast models
Smoothing methods, such as moving averages and exponential smoothing are widely employed for forecasting purposes in many production and operations management applications, largely because of their simplicity and ease of implementation. As such, most researchers of supply-chain management (SCM) problem requiring a forecast model have based their studies on either the moving-average method (e.g., Chen et al., 2000a) or the exponential weighted moving-average (EWMA) method (e.g., Chen et al., 2000b). Given the close connection between the moving-average method and the EWMA method, we will only focus on one method (namely, the EWMA method) in this paper. The EWMA model can be expressed as follows: Ft+1|t = dt + (1- ) Ft|t-1
Page 13
where 0 < < 1 is the smoothing constant and Ft+1jt is the forecast of period t + 1 made at the end of period t. It should be noted that the forecasts for periods t + i (i = 1,2,. . .) made at time t are equal, that is, Ft+i|t = Ft+1|t for i = 1,2,. . .. Hence, the forecasts for all lead times will follow a horizontal line parallel to the time axis.
Even though the EWMA method, and to a lesser extent the moving-average method, has flexibility for adapting to a variety of correlated demand processes, it is MSE optimal for only one underlying time-series model, namely, a first-order integrated moving average, denoted by ARIMA(0, 1, 1) or IMA(1, 1) (e.g., Graves, 1999). An ARIMA (0, 1, 1) process is a nonstationary process that can be interpreted as a random walk trend plus a random deviation from the trend. Thus, under no circumstance is the EWMA method MSE optimal for a stationary AR(1) process. This fact opens up consideration of employing an MSEoptimal forecast scheme for the assumed AR(1) process. By recursively applying (1), it is easy to show that:
(2) For a general ARIMA process, it can be shown that the minimum mean square error forecast for period t + i is the conditional expectation of dt+i given current and previous observations dt, dt-1, dt-2,. . . (see Box, Jenkins, & Reinsel, 1994). In the case of an AR(1) process, this implies the MSE-optimal forecast function is given by E(dt+i|dt). Since E(t+i|dt) = 0 for i = 1,2,. . ., it immediately follows that for an AR(1) process, the MSE-optimal forecast function is given by:
|
1 1
In contrast to the two previous methods, this forecast function is not a horizontal projection into the future. Instead the forecasts revert back towards the overall mean level of /1-. The MSE-optimal forecast function reflects the fact that the AR(1) process is stationary and has the property of conditional mean reversion; that is, even though the process can be expected to wander away (below or above) from the overall mean it is also expected to eventually return back to the overall mean. The moving-average and EWMA methods fail to capture this mean reversion property of a stationary AR(1) process.
One argument often presented against the use of optimal methods is that their implementation is more difficult than the simple smoothing methods when parameters are unknown. It is pointed out that to implement optimal methods requires statistical skills in time-series modelling, including knowledge of model identification, model estimation, and tests for model adequacy, that are beyond the skill set of a typical operations manager. However, we believe that the industrial use of more sophisticated time-series models is steadily growing because of two reasons. First, the requirement of intense statistical training, often referred to as 6 training, is increasingly becoming commonplace (Hoerl, 1998). At corporations like GE, Motorola, and Allied Signal, organizational cultures are being developed in which there is a strong desire from employees throughout the organization to learn and implement advanced statistical techniques. Indeed, the authors of this paper can report that seminars in time-series analysis are part of the regular continuing education program at GE-Medical Systems and are required to be taken by all supply-chain managers. Second, modern computational tools are readily available to make possible automated implementation of time-series modelling including the general class of ARIMA models. These programs are designed to automate model identification, model fitting, and forecasting.
Page 15
CHAPTER 5
5. Analysis done on each data set
Each data set was carefully analysed and the above mentioned properties were checked for. The data collected for each crop or the yield of each crop over the last 25 years had the characteristics that were required for the model to be applied and tested. All analysis done on each data set is being listed below and some of them are being explained in detail. Check for Auto-Correlation. Calculation of . Calculation of the optimal period for moving average Calculation of . Calculation of the errors. Checking the normal behavior of the errors. Forecasting using both models. Comparison of the results.
Calculation of and
As we have already mentioned above that the demand is assumed to follow the first order autocorrelation. In a first order auto correlated demand, the demand is being related to the previous year demand as given by the equation:
When we closely observe the relationship, we can see that this follows a regression, where the independent variable is the present demand and the dependent variable is the previous period demand. Thus, applying the basic concept in the determination of the parameters for the regression analysis, we assume that for the best fit line the error is zero. This leaves us with the equation for the straight line. For this straight line we can see that the slope of the line, in this case which is given by can be calculated using the equation
( (
Page 16
Where d is the two period moving average. Here we have taken the two period moving average other than the average over the period, since with two period moving average we have found that the error in the forecasting was comparatively lesser than that with the average over that entire data. On further thought why is this phenomenon being observed we concluded that this could be attributed to the reverting property of the AR(1) models. This property tends to bring the forecast closer to the mean. But for a positively correlated data this leads to a large error. This drawback can be solved by going for a moving average. The period for the moving average was determined using spreadsheet simulation, where 2,3,4 and 5 period moving average was tested. It was surprisingly found that for all data sets 2 period moving average yielded the best results. This can be explained by the fact that the factors affecting the agriculture do not change abruptly. They follow a gradual change. The changes could be technological, physical (government policies) or even natural. Thus the forecast yielded best results with 2 period moving average. Statistically speaking, the data showed large amount of first order autocorrelation. And for higher orders it became less significant.
Page 17
CHAPTER 6
6. Results
The above analysis will be explained with the help of a single data set. This data set that we have selected is for rice and paddy. The test static obtained for Durbin-Watson test was 1.964. since the value is less than 2 the data can be said to be first order auto correlated. The value of was obtained to be 0.965 and as 0.915. The normality of the error was analysed with the help of the statistical packages like SPSS and Minitab. The mean and standard deviation was found to be 7.2 x 10-16 and 0.978. this shows that the mean is almost zero and the normality of the error was checked with the help of the probability value and the normal probability plot. The p value was found to be 0.583. Since the probability value is less than the value of critical value of 0.05 at 95% confidence level the hypothesis that the error follows the normal distribution can be accepted. The plots are being shown below.
Probability Plot of C1
Normal
99 Mean StDev N AD P-Value 0.0000004167 0.9780 24 0.290 0.583
95 90 80 70
Percent
60 50 40 30 20 10 5
-3
-2
-1
0 C1
Figure 2 : Histogram and normal probability plot for rice and paddy
The fore casting using the two methods has been shown in the table given below. It can be seen that MSE optimal gives the better forecast for the agricultural yield.
Page 18
Table 2: Forecast results for rice and paddy

year procduction in MT moving avg =(1-) forecast using MSE error in forecast error in forecast using forecast for MSE EWMA for EWMA 2.6E+12 5.1E+11 1.6E+12 1.6E+13 1.3E+12 1.1E+12 4.7E+11 1.8E+09 7.6E+12 2.2E+12 1.1E+11 4.5E+12 1.9E+12 5.2E+12 7.3E+12 3.3E+11 1.6E+13 1.7E+13 2.1E+13 5.3E+10 1.4E+13 7.2E+12 1.3E+13 1.5E+13 5.4E+12 1.6E+07 1.9E+07 1.8E+07 1.7E+07 2.1E+07 2.2E+07 2.3E+07 2.3E+07 2.2E+07 2.4E+07 2.5E+07 2.4E+07 2.5E+07 2.5E+07 2.6E+07 2.7E+07 2.6E+07 2.8E+07 2.2E+07 2.7E+07 2.5E+07 2.8E+07 2.8E+07 2.9E+07 9.8E+12 2.5E+11 1.8E+12 1.6E+13 1.3E+12 5.3E+11 3.4E+10 4E+11 5.4E+12 4.4E+11 2.1E+12 1.8E+12 1.2E+11 1.3E+12 1.5E+12 1.9E+12 6.4E+12 4.3E+13 2.2E+13 1.7E+12 6.7E+12 2.7E+11 1.4E+12 7.6E+11 6.3E+12
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
17427944 19081147 18314686 17013045 21171447 21969341 22597604 22721384 22074476 24449500 24917312 23404729 24877437 25106636 26218740 27361764 25890853 28546322 21808324 27050186 25332886 28042829 28341887 29476025 30246312
18254546 18274592 17959206 18601654 19162935 19653602 20037075 20263453 20682057 21067081 21261885 21540004 21794763 22089695 22419200 22623414 22952465 22892247 23100144 23206465 23426300 23640021 23883188 24137713
0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653 0.9653
633370 634065 623123 645413 664888 681912 695217 703072 717596 730955 737714 747364 756203 766436 777869 784955 796371 794282 801495 805184 812812 820227 828664 837495
17456624 19026149 18278953 17222310 20846360 21527184 22036830 22116475 21695373 23443188 23740679 22760626 23742623 23929877 24656318 25316492 24564575 25886655 22439731 25103297 24334533 25664599 25862329 26425083 mean square error
The results of the analysis on the other crops are given in the table below.
Page 19
Table 3: Results for the forecasting of other crops

crop durbinwatson test static 1.964 0.764 0.569 0.814 1.138 1.001 1.132 1.389 1.556 0.939 error in forecast for MSE 5.446E+12 1.405E+11 2.58E+12 2.577E+12 4.429E+11 6.922E+10 1.487E+11 1.621E+10 3.21E+11 1.663E+09 error in forecast for EWMA 6.3E+12 2.6E+11 3E+12 3.8E+12 7E+11 1.6E+11 2E+11 2.6E+10 3.7E+11 2.9E+09 p- value for the error 0.583 0.005 0.139 0.005 0.02 0.305 0.457 0.148 0.707 0.648
rice and paddy sugarcane wheat buffalo milk vegetables potato millet maize groundnuts coconut
0.965 0.961 0.977 0.930 0.968 0.906 0.967 0.938 0.974 0.957
0.915 0.904 0.915 0.833 0.923 0.778 0.919 0.852 0.936 0.896
From the table it can be seen that the buffalo milk, sugarcane and the vegetables do have the errors which follow the normal distribution. In this case the basic assumption is being violated and so the data do not follow the demand distribution equation. Thus they have to be considered separately and the other forecasting techniques should be used for their prediction.
Page 20
Figure 3 : Histogram and normal probability plot for sugarcane
Figure 4 : Histogram and normal probability plot for wheat
Page 21
Figure 5 : Histogram and normal probability plot for milk
Figure 6 : Histogram and normal probability plot for vegetables
Page 22
Figure 7 : Histogram and normal probability plot for potato
Figure 8 : Histogram and normal probability plot for millet
Page 23
Figure 9 : Histogram and normal probability plot for maize
Figure 10 : Histogram and normal probability plot for groundnut
Page 24
Figure 10 : Histogram and normal probability plot for coconut
Page 25
CHAPTER 7
Conclusion
In this research, we found that in most of the cases the agricultural yield tends to follow the first order auto correlation. This property was tried to exploit and by using MSE optimal forecasting model the agricultural produce for next period was tried to be forecasted. The MSE optimal model was compared to the exponential weighted mean average method of forecasting and it was found that the MSE optimal was gave a better estimate. The performance parameter that was taken for comparison was the mean square error. The MSE optimal method gave forecasts with lesser amount of mean square error. The normality of the errors were also studied and expect for a few all other gave the results that the errors were normally distributed and they had a mean of zero. In practice though the mean was not actually zero but the value was small enough to be approximated to zero. In this research we had restricted ourselves to the analysis were we tried to capture only the effect of the previous year yield. But in actual practice it is not so. Agriculture depends upon a lot of factors such as weather, new technological development, new variety of seeds, government policies, etc. since some of the factors which affect the agriculture are not quantitative it becomes all the more difficult to capture their effect. However this opens a wide area for future research.
Page 26
References Alwan, L. C. (2000). Statistical process analysis. Burr Ridge, IL: McGraw Hill/Irwin. Alwan, L. C., Liu, J. J., & Yao, D. Q. (2003). Stochastic characterization of upstream demand processes in a supply chain. IIE Transactions, 35, 207219. Biggs, J. R., & Campion, W. M. (1982). The effect and cost of forecast error bias for multiple-stage production-inventory systems. Decision Sciences, 13(4), 570584. Box, G. E., Jenkins, G. M., & Reinsel, G. C. (1994). Time series analysis: Forecasting and control (3rd ed.). Englewood Cliffs, NJ: Prentice Hall. Chen, F. Z., Drezner, Z., Ryan, J. K., & Simchi-Levi, D. (2000a). Quantifying the bullwhip effect in a simple supply chain: The impact of forecasting, lead times, and information. Management Science, 46, 436443. Chen, F., Ryan, J. K., & Simchi-Levi, D. (2000b). The impact of exponential smoothing forecasts on the bullwhip effect. Naval Research Logistics, 47, 269286. Enns, S. T. (2002). MRP performance effects due to forecast bias and demand uncertainty. European Journal of Operational Research, 138, 87102. Graves, S. C. (1999). A single-item inventory model for a nonstationary demand process. Manufacturing and Service Operations Management, 1, 5061. Graves, S. C., &Willems, S. P. (2000). Optimizing strategic safety stock placement in supply chains. Manufacturing and Service Operations Management, 2, 6883. Hoerl, R. W. (1998). Six sigma and the future of the quality profession. Quality Progress, 3542. Hopp, W. J., & Spearman, M. L. (2000). Factory physics (2nd ed.). Boston, MA: McGraw-Hill/Irwin. Jonsson, P., & Mattsson, S.-A. (2003). The implications of fit between planning environments and manufacturing planning and control methods. International Journal of Operations & Production Management, 23, 872900. Kahn, J. (1987). Inventories and the volatility of production. The American Economic Review, 77, 667679. Lee, T. S., & Adam, E. E. (1986). Forecasting error evaluation in material requirements planning (MRP) production-inventory systems. Management Science, 32(9), 11861205.
Page 27
Lee, H. L., Padmanabhan, P., & Whang, S. (1997). Information distortion in a supply chain: The bullwhip effect. Management Science, 43,546558.
Lee, H. L., So, K. C., & Tang, C. S. (2000). The value of information sharing in a two-level supply chain. Management Science, 46,626643.
Page 28

Comparison of Forecasting Methods For Agriculture

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Comparison of Forecasting Methods For Agriculture

Enviado por

Direitos autorais:

Formatos disponíveis

Comparison of forecasting methods for agriculture