Issues Magazine

Seasonal Climate Forecasts: Can Decision Analysis Help Agriculture?

By Peter Hayman

Because decisions are about the future they involve uncertainty, yet we make most decisions with a minimum of effort. Hard decisions usually involve high levels of uncertainty and have significant consequences: they are risky decisions. Climate is a major source of uncertainty, often with significant consequences for agricultural decision-makers.

The cost of uncertainty to agricultural decision-makers is twofold and can be summarised as the moving target effect and the protective strategy effect (F.J. Meza, J.W. Hansen and D. Osgood, J. Appl. Meteorol. Climatol. 2008, 47:1269–86 2008). The moving target effect is where decisions such as crop type or rate of fertiliser have to be made prior to the season unfolding. Climate variability almost always leads farmers to use too little or too much of an input. If aiming for an average, in some years they will miss by a small margin and in other years by a considerable margin.

While the moving target effect is a cost to all farmers irrespective of their risk preference, the protective strategy effect is a cost to risk-averse farmers who use suboptimal inputs and therefore do not capitalise on favourable seasons. There may also be environmental costs whereby protective strategies involve practices such as leaving a field bare to store water. This will increase production in a drier-than-average season but increase the risk of erosion in wetter seasons.

Climate science has two related types of information to offer agricultural decision-makers dealing with climate-related risk. First, making available the historical climatological data allows a decision-maker to answer the “What if?” questions along the lines of “If I were to apply this amount of fertiliser, how often would I have made a profit over the historic record?” and the related “What’s best?” question, for example “What was the optimum rate of fertiliser to apply over the historic record?”. In many cases these historic databases can be used in a simulation model or a simple statistical relationship between yield and rainfall. The main advantage of accessing historical climate records is to extend the knowledge base of the farmer.

The second source of information is a forecast of the coming season. Although this can be presented categorically as “the coming season will be drier than normal”, this information is best presented as a shift in probabilities from the historical climatological odds of 50/50: “There is a 70% chance that the coming season will be drier than normal and a 30% chance that it will be wetter”.

The need for seasonal climate forecasts to assist agricultural decision-making in one sense is as old as agriculture. Recent high input costs, food shortages, concerns over agricultural impacts on the environment and projected climate change have increased the call for climate information to be made available to agricultural decision-makers.

An underlying assumption is that, as a discipline, climate science has information that is useful to agricultural decision-makers, and that more effort is required in communication and dialogue. This assumption is reasonable and can be supported by individual success stories ranging from smallholder farmers in developing countries to large commercial farms in the US, Europe, South America and Australia.

However, mixed with the success stories is a recognition that the task of communicating information to improve the management of weather and climate risk in agriculture is more difficult than first thought.

Being Clear about Communication
The simplest interpretation of the challenge to develop a better pipeline between the information provider (agrometeorological services) and information user (agricultural decision-makers) is by building software, using smarter hardware or writing clearer bulletins. This interpretation is based on a notion that knowledge is created by research, communicated by intermediaries, and used by farmers. It also implies that knowledge can be transferred as an unambiguous signal. While this may be appropriate for some technologies, it is inappropriate for communicating climate risk. The adoption of embodied technologies such as a new crop variety or crop protection chemical is inherently simpler than information-based technology such as a seasonal climate forecast. Not only is risk and decision-making complex, the communication challenge is greater because the knowledge about managing climate risk lies as much with the farming community as it does with the discipline of agricultural science or climate science.

Defining Risk
Although the notion of risk and managing risk is so commonplace nowadays as to become a cliché, it is a relatively recent phenomenon. Anthony Giddens, in Runaway World: How Globalisation Is Reshaping Our Lives (2002), maintains that the idea of risk only took hold in Europe in the 16th and 17th centuries, where the term came to the English language through Spanish or Portuguese, referring to sailing into uncharted waters with the chance of great gain weighed against the chance of loss. According to Peter Bernstein in Against the Gods: The Remarkable Story of Risk (1996), the term “risk” was derived from the Italian riscare, which means “to dare”. He linked risk and modernity:

The revolutionary idea that defines the boundary between modern times and the past is the mastery of risk: the notion that the future is more than a whim of the gods and that men and women are not passive before nature. Until human beings discovered a way across that boundary, the future was a mirror of the past or the murky domain of oracles and soothsayers who held a monopoly over knowledge of anticipated events.

Risk may be part of modernity but the discussion of risk regarding pollution, food safety and climate change is seldom free from emotion. Psychological studies have identified various issues that influence the perception of risk including the subject’s sense of control and world view, whether a risk is voluntary, and the distribution of costs and benefits. The matter is further complicated by the way society finds some risks acceptable while holds other risks as special (e.g. the very different treatment of equal probabilities in car accidents versus aeroplane accidents). Risk can only be understood in the social and psychological context of the decision-maker.

Apart from the social context of risk, communication is further hampered because few of us are intuitive statisticians. When faced with uncertainty we rely on mental shortcuts that can be efficient but sometimes lead to biases that impair the decision-making process. As noted by Paul Slovic in Facts Vs Fears: Understanding Perceived Risks (1984), “it is extremely hard for people to think about uncertainty, probability and risk.”

Communicating Skilful but Uncertain Forecasts for Risky Decisions
It is common for intermediaries such as agronomists to state that while it might suit climate science to use probabilities, farmers need a categorical forecast “because in the end they need to make a decision”. Implicit in this statement is the notion that probabilistic forecasts can’t be used in decision-making.

There is little difficulty in knowing what to do with a very accurate categorical forecast as it fits the easy logic of IF, THEN, ELSE. IF the season ahead is going to be a drought, THEN destock heavily, ELSE continue as normal. As pointed out by Nassim Nicholas Taleb in Fooled by Randomness (2004), the issue is deeper than struggling with the mathematical aspects of probability (which are very simple for decision analysis). He reviewed the large amount of psychology literature that shows that most of us are probability blind: we struggle to think of alternative futures (much less alternative histories).

Kevin Parton’s paper “Agricultural Decision Analysis: The Causal Challenge” (Australian Agricultural and Resource Economics Society Conference, 2009) reviewed studies showing that business decision-makers rarely approach a problem with a probabilistic mindset, but more commonly have a causal mindset. He identified three approaches that relied on causal intuitive thinking. The first is scenario construction whereby a future path is chosen, committed to by the decision-maker, and then effort is made to “make the decision work”. For example, many farmers have a plan for their farm and are reluctant to shift from that plan.

The second is role-based considerations whereby the decision-maker recognises a situation where the best action is known. An example is when the rains come late, a farmer responds to a series of rules for appropriate crops in late sowing and does not weigh up risks or use forecasts.

The third approach is relying on social norms. For example, some (not all) farmers were reluctant to sow at a different time or a different crop from their neighbours because failure might lead to social ridicule.

Following his review of the literature and after conducting interviews with Australian grain farmers who had been exposed to seasonal climate forecasts and simulation modelling, Parton noted that if rule-based planning based on causal thinking is the norm, this is a significant challenge to the use of probabilistic information.

Although they present difficulties for communication and use, there are a number of good reasons to present forecasts as probabilities: first, because it is a more accurate statement of the understanding from climate science; and second, because it encourages intermediaries (e.g. agronomists) and farmers to practise risk management.

Pierre de Laplace (1749–1827) stated that probability has reference partly to our ignorance and partly to our knowledge. The atmosphere is a complex chaotic fluid, and although patterns of ocean temperatures “nudge” this chaos in certain directions there will always be a significant proportion of unexplained variation.

Indeed, along with increased understanding of the climate–atmosphere system has come a better understanding of theoretical and practical limits to prediction. As a means of conveying this, one approach is to use a spinning probability disc divided into thirds to represent the chance of falling into each tercile (the lowest, mid and highest third of the data values). Seasonal climate forecasts are shown as shifts in the three sections of the pie chart. The idea is to convey the combination of knowledge (change pattern on disk) and ignorance (exactly where the disk will stop spinning).

Will Coventry’s Honours thesis at the University of Queensland in 2000 suggested that some of the problems communicating probabilities could be overcome by expressing the probability as a frequency (for example, out of 10 times that the sea surface temperatures were in this pattern, seven of them were wetter than median and three were drier). The reasoning is that when hearing about a 70% probability of being wetter, people are less likely to appreciate the 30% chance of being drier. The Australian Bureau of Meteorology use frequencies to explain its forecasts.

A second reason for probabilistic forecasts is to ensure that they improve rather than hinder risk management. The perversity of seasonal climate forecasts is that if they are misunderstood as categorical forecasts, they can lead to poorer risk management than if the farmer had never heard of the forecast. In the absence of a forecast, a farmer may plan for a wide range of outcomes. However, if this is adjusted to one outcome in the mind of the decision-maker, the forecast has clearly been misleading in terms of risk management.

Does Decision Analysis Have a Place?
The probability disk is one of a number of ways to convey the uncertainty in the forecast. In some cases, when farmers appreciate the level of uncertainty they decide not to use the forecast and concentrate on those aspects of the farming system that they can control. Others pose the challenge: how do you use this form of uncertain information for decision-making?

One approach, decision analysis, provides a logical framework for a decision-maker to formulate preferences, assess uncertainty and make judgements. Although agricultural economists have long used decision analysis, its use by agrometeorologists and agricultural scientists has been limited.

There has been a tradition in agricultural science to talk the language of choice–consequence. For example, if you put on x units of nitrogen, you will get a yield of y. This is fine for reviewing experiments; however, looking forward we need to consider the language of choice–chance–consequences. In other words, if you put on x units of nitrogen, depending on the season type you will get a yield of y1, y2 or y3.

An Example of Nitrogen Management on Wheat
A farmer deciding on the rate of nitrogen fertiliser has to balance potential demand from the crop with the supply from the soil and fertiliser. This balancing act is made more difficult when dealing with uncertain crop demand. What is the best rate of fertiliser to use in the absence of seasonal climate forecasts? How can uncertain seasonal climate forecasts be used to change the decision?

1. Define the choices with single consequences: Although there is an infinite range of fertiliser rates, the options can be simplified as fertilising for a low demand in a poor season (20 kg of nitrogen per hectare in the bottom third of years), an average season (60 kg of nitrogen per hectare in the middle third of years) or a good season (100 kg of nitrogen in the best third of years). Figure 1 is a simplified decision tree that shows the outcome as gross margin per hectare for 20 kg, 60kg and 100 kg of nitrogen. This tree is an oversimplification but it shows what we would do with perfect seasonal climate forecasts or what would have been best in hindsight. It also shows that the payoff for nitrogen is asymmetrical, and there are very good returns in good seasons.

2. Define a range of consequences for each choice: While Figure 1 shows a single consequence for each choice, it avoids the point that any fertiliser rate is likely to overfertilise or underfertilise the crop depending on what type of season actually eventuates. Rather than looking at the three branches of Figure 1, a more complete picture is shown by the nine branches in Figure 2. When a farmer aims low, the result will either be a balanced budget in a poor season or a nitrogen limit in an average or good season. The other extreme of aiming high will result in a balanced budget in a good season and water-limited but excess nitrogen in average or poor seasons.

When the gross margins for each of the nine outcomes are considered as shown in Figure 3, it is clear that fertilising for a poor season is the low-risk, low-return option while fertilising for a good season has high potential returns but also a chance of a negative gross margin. We have made the initial assumption that the chances of a poor, average or good season are equal (as indicated by the pie chart), and the probability-weighted average reflects this. The probability-weighted average or expected value never actually occurs; rather, it is a summary statistic.

Before the season unfolds, farmers must choose one of the strategies: “fertilise for poor season”, “fertilise for average season” or “fertilise for good season”. A risk-averse farmer could quite rationally select the first of these to avoid the risk of poor outcomes that is a component of the other strategies. In a less risk-averse manner another farmer could select the average fertiliser rate and be prepared to sacrifice the difference in the probability-weighted average to avoid the 33% chance of making a loss that is contained in the “fertilise for good season” strategy.

Figure 4 shows a seasonal forecast in which there is an increased probability of a poor season. Seasonal forecasts do not change the outcomes; they just change the chances of the outcomes and therefore the probability-weighted average. In this case, even though there is a 55% chance of a poor season, because of the asymmetry of pay-offs for nitrogen, the expected value is still maximised by fertilising for an average season. The impact of such a forecast across the farming community should be to shift farmers towards lower fertiliser application rates, because the risk of poor outcomes at higher rates of fertiliser application has increased.

In our experience, presenting this as a spreadsheet to farmers and agronomists is an effective way to start a dialogue on how uncertain forecasts might be used to change fertiliser decisions. It also reinforces the need to consider the minority outcome (a poor season when the forecast is for a good season). In the case of nitrogen, due to the good pay-offs in the good seasons it indicates that the forecast has to be quite negative to reduce rates.

Modified from P. Hayman, “Communicating Climate Risk: Choices, Chances and Chocolate Wheels”. Presentation at Greenhouse 2011 conference, Cairns, April 2011.