Instead of giving a single number forecast, the India Meteorological Department must switch to a probabilistic approach
Given the current rainfall trends in June and July, the monsoon (June 1-September 30) rainfall for the country as a whole is in all likelihood to be “deficient” (defined as less than 90 per cent of LPA, the Long Period Average). The shortfall in July is 22 per cent, which is very unlikely to be made up by the rainfall in the remaining two months and prevent the imminent (meteorological) drought.
From public perception, the India Meteorological Department (IMD) failed in its forecast for which it is already drawing a lot of flak from media commentators for getting it wrong “once again.” The IMD had forecast a ‘normal’ monsoon with a total seasonal rainfall for the country as a whole at 96 / 4 per cent of the LPA. But, actually, the IMD forecast this year was pretty good and this flak unwarranted. Yes, this one sentence forecast aimed at giving a single number for the consumption of the public and politicians is definitely off the mark. But there are other significant details in the IMD’s changed and improved forecasting methodology that it has been following since 2003, which would show the IMD in better light.
The new strategy has been to move away from a deterministic forecast of this single quantity, the India Summer Monsoon Rainfall (ISMR), to a probabilistic forecast. In fact, India is the only country which gives a quantitative long-range forecast (LRF) when long-range forecast, whether statistical or dynamical, is a highly probabilistic exercise depending very critically on the initial (summer-time) values of the meteorological variables and the models used. In the statistical approach, the variables are certain meteorological (regional and global) parameters that have been found to be statistically well-correlated to the ISMR. In the dynamical approach they are the values of physical variables themselves, such as pressure, wind velocity, etc.
A statistical approach is resorted to because the underlying physics of the monsoon is not yet fully understood for the dynamical equations to correctly represent the causative conditions for the monsoon and its evolution. Moreover, integrating these highly non-linear equations over a long period for an LRF requires high-resolution data on the variables (not easily available) and highly intensive computational exercise requiring huge high-performance computing resources. Also, because of non-linearity, small (measurement) errors in the initial conditions will diverge very quickly to yield wrong results. In fact, to date, no dynamical model has been able to simulate the monsoon over the Indian region well and accurately predict the monsoon behaviour. For example, the model from the National Centers for Environmental Prediction (NCEP) of the United States, which has been adapted for monsoon prediction by the Indian Institute of Tropical Meteorology (IITM), Pune, has forecast 104 per cent rainfall for 2012.
The statistical approach, on the other hand, depends on how robust and stable the atmospheric forcing parameters (the predictors) are. A major problem has been to identify a small set of stable and independent parameters that influences the monsoon rainfall and the bulk of its variance. In fact, there is a natural variability of the rainfall on the decadal scale that is seen from historical data. A 30-year moving average plot of the rainfall suggests that we are currently in the “below normal” epoch of this natural variability. Moreover, the predictors themselves have been found to be unstable over long periods. Many of the once strongly influencing parameters have declined in their correlations over the years. Some have, in fact, turned negative. And yet the monsoon system itself seems to be stable and has been visiting us every year without fail! The search for a minimal set of stable and strongly enforcing parameters thus remains a constant one.
Therefore, instead of a single-number-fixated forecasting exercise, probabilistic forecast makes eminent sense and logic. Also, since the forecast skill of individual models has been found to be not very good, since 2007, the IMD has adopted an ensemble approach. Here the ensemble includes statistical models of its own and different dynamical models from various international organisations. This approach will also give probabilities, rather than a single definitive number, for different outcomes of the ISMR.
For this probabilistic exercise, the IMD has classified the monsoon rainfall into five categories: Deficient (less than 90 per cent of the LPA), Below Normal (90-96 per cent), Near Normal (96-104 per cent of the LPA), Above Normal (104-110 per cent) and Excess (above 110 per cent). In any statistical exercise, there will be a finite probability of the outcome being in any of these categories. The confidence with which a prediction is made depends on how well one is able to estimate these relative probabilities.
But before looking at the IMD’s 2012 forecast, let us see what the a priori (purely climatological) probabilities for these categories are. The climatological probability for a “Normal” monsoon (96-104 per cent) in any given year is 33 per cent and those for “Below Normal” and “Deficient” monsoon are 17 and 16 per cent respectively, which are by no means insignificant. These numbers need to be kept in mind when one looks at the actual forecast and see how much the year’s meteorological conditions change these a priori probabilities.
Unfortunately, the IMD has not felt it necessary to emphasise this aspect in its forecast. Apart from 2003 and 2004, when this probabilistic approach began, and more recently in 2011-2012, information on the relative probabilities for the five categories was not made public. Apart from political expediency and possible adverse influence on the market, one cannot imagine any other reason. Given the convenience of pegging stories to this single number, news reports too have ignored the other details of this year’s forecast. So the IMD itself is partly responsible for inviting criticism after this year’s forecast.
So what did this year’s forecast tell us in terms of probabilities? In the preliminary April 26 forecast, which is based on December-March data of atmospheric variables, the IMD forecast was 47 per cent probability for a “Normal” monsoon, 24 per cent for a “Below Normal” monsoon and eight per cent for a “Deficient” monsoon. In the updated June 22 forecast, which includes data up to May, a possible trend towards a poor monsoon was discernible. As compared to the April forecast, while the probability for a ‘Normal’ monsoon fell to 42 per cent, “Below Normal” shot up to 35 per cent and “Deficient” increased to15 per cent, nearly the climatological probability. Unfortunately, the IMD does not sufficiently emphasise these aspects in its press release and continues to give undue focus on that single number in spite of its changed forecast strategy. It is clear that these probabilities give sufficient insight to possible monsoon behaviour and can serve as guidance for proper planning.
But it is not clear in what form the forecast is presented to the planners and the agriculture ministry for them to take appropriate measures because the way the government’s response is being projected in media reports it would seem that it did not have any idea of the distinct possibility of a bad monsoon year. It can even be argued that the IMD totally gives up this single number forecast for the ISMR. Instead, different agencies can scientifically interpret the forecast probabilities for the different categories and take appropriate contingency measures. If this information is to be continued to be shared with the public, an exercise that began only in 1988, then the IMD must also take efforts to explain the probabilistic nuances of the forecast to the media and the public. It is high time that the IMD moved away from realpolitik to real-scientific.