Sustainability & Environment (Agri-Food Systems) · Climate Science & Agriculture · also Climate Modelling, Climate Projections and Scenario Analysis, Climate Information Services
How future climate is projected and how to use projections responsibly: model hierarchy, CMIP ensembles, RCP and SSP scenarios, downscaling and bias correction, and uncertainty. Needed for impact studies, climate-risk disclosure and adaptation planning.
In the exams: GARP SCR 'Climate Models and Scenario Analysis'; PG and PhD research in agrometeorology and climate change; ICAR AICE subject 60 (climate variability)
Climate models and scenarios are how future climate is projected, and downscaling and bias correction are how projections are made usable at the scale of a farm, district or company asset. This subject covers the hierarchy of models from simple energy-balance models to Earth system models and regional models; the Coupled Model Intercomparison Project (CMIP) and how models are evaluated; the scenario families (SRES, RCPs, SSPs) and the integrated assessment models that produce mitigation pathways; statistical and dynamical downscaling; bias correction; the sources of uncertainty; seasonal and sub-seasonal forecasting; climate and extreme-event indices; the open datasets and portals practitioners use; and scenario analysis for organisations, as required by climate-risk disclosure standards.
It sits between two companion subjects. Climate System Science supplies the physics (forcing, feedbacks, sensitivity, carbon budgets) that models represent; Climate Change Impacts on Agriculture & Food Systems uses model output to drive crop models and impact assessments. This manual does not repeat the physics or the crop modelling; it concentrates on how projections are produced, what they can and cannot say, and how to use them responsibly.
The subject is examined in research and professional settings more than in entrance papers. The GARP Sustainability and Climate Risk certificate has a full section on climate models and scenario analysis, including SSPs, IAMs and NGFS scenarios. ICAR AICE Agricultural Meteorology (subject 60) and CSIR NET atmospheric science ask about model types, ENSO and monsoon forecasting, downscaling and indices. PG and PhD researchers in agrometeorology and climate change need the practical skills of finding data, downscaling, bias correcting and quantifying uncertainty. Corporate and financial practitioners need to run and explain scenario analysis under TCFD-aligned and ISSB disclosure rules.
Climate science uses a hierarchy of models, each suited to particular questions. The simplest is the zero-dimensional energy-balance model (EBM), which treats the Earth as a single point: C dT/dt = F − λT, where C is the effective heat capacity, F the radiative forcing, λ the climate feedback parameter and T the temperature anomaly. At equilibrium T = F/λ, and the response time is τ = C/λ. With C set by a 100 m ocean mixed layer (about 4 × 10⁸ J m⁻² K⁻¹) and λ = 1.2 W m⁻² K⁻¹, τ is about 11 years. Two-layer EBMs add a deep ocean and reproduce the fast and slow responses of complex models; they are used as emulators in IPCC assessments and in integrated assessment models. One-dimensional EBMs resolving latitude (Budyko and Sellers, 1969) showed how ice-albedo feedback can produce multiple climate states. Radiative–convective models, pioneered by Manabe and Wetherald (1967), resolve the vertical column and gave the first physically sound estimate of warming from doubled carbon dioxide; Syukuro Manabe shared the 2021 Nobel Prize in Physics for this line of work. General circulation models (GCMs) solve the equations of fluid motion, thermodynamics and radiative transfer on a three-dimensional grid for the atmosphere and ocean, coupled with land and sea-ice models (atmosphere–ocean GCMs, AOGCMs). Earth system models (ESMs) add biogeochemical cycles, especially the carbon cycle, so that they can be driven by emissions rather than prescribed concentrations and can simulate carbon–climate feedbacks; some add atmospheric chemistry, aerosols and dynamic vegetation. Earth system models of intermediate complexity (EMICs) simplify the dynamics to run for millennia. Typical CMIP6 global models have atmospheric grid spacing of about 100–250 km.
Processes smaller than a grid box, such as individual convective clouds, turbulence, cloud microphysics and land-surface heterogeneity, cannot be resolved and are represented by parameterisations: simplified relations that express their aggregate effect in terms of resolved variables. Parameterisations contain tunable parameters, and differences between them, especially for clouds and convection, are a main reason why models differ in climate sensitivity and regional rainfall. Model tuning adjusts parameters within plausible ranges so that a model reproduces key observed quantities, such as the top-of-atmosphere energy balance; documenting tuning is now expected practice. Regional climate models (RCMs) run at higher resolution, typically 10–50 km, over a limited area, taking their boundary conditions from a GCM or a reanalysis. They resolve topography, coastlines and land-use contrasts better and can add value for rainfall in mountains and monsoon regions, but they inherit the large-scale errors of the driving GCM. The Coordinated Regional Climate Downscaling Experiment (CORDEX) organises RCM simulations for defined domains such as South Asia, Africa, Europe and others. Convection-permitting models with grid spacing of about 4 km or less resolve deep convection explicitly and better simulate hourly rainfall extremes and the diurnal cycle, at great computational cost. Choosing a model is choosing a trade-off between process detail, resolution, ensemble size and computing cost. A large ensemble of a simple model can explore uncertainty that a single high-resolution run cannot, while a high-resolution run can represent processes a simple model omits. Good practice is to use the simplest model adequate to the question.
The Coupled Model Intercomparison Project (CMIP) is organised under the World Climate Research Programme's Working Group on Coupled Modelling. It defines common experiments, forcing data and output formats so that results from modelling centres around the world can be compared and combined. CMIP3 underpinned the IPCC Fourth Assessment Report, CMIP5 the Fifth and CMIP6 the Sixth; CMIP7 is under way. The core of CMIP6 is the DECK (Diagnostic, Evaluation and Characterization of Klima): a pre-industrial control run (piControl), an atmosphere-only run with observed sea surface temperatures (AMIP), an abrupt quadrupling of carbon dioxide (abrupt-4xCO2, used to estimate ECS) and a 1 per cent per year carbon dioxide increase (1pctCO2, used for TCR), plus a historical simulation from 1850. Endorsed model intercomparison projects (MIPs) add specialised experiments; ScenarioMIP runs the SSP-based future scenarios, and others cover detection and attribution, land use, clouds, the carbon cycle and more. A multi-model ensemble samples structural uncertainty: different reasonable ways of building a model. The ensemble mean often matches observed climatology better than most individual models because errors partly cancel. But the ensemble is an 'ensemble of opportunity', not a designed random sample: models share code, parameterisations and tuning targets, so they are not independent, and the spread does not necessarily cover the true uncertainty.
Model evaluation compares simulations with observations and reanalyses for the mean climate, seasonal cycle, variability (such as ENSO and the monsoon), trends and extremes. Common statistics include bias, root mean square error, correlation of spatial patterns and the Taylor diagram, which shows on one plot a model's pattern correlation, its standard deviation relative to observations and its centred RMS difference. Process-oriented evaluation checks whether a model gets results for the right reasons, for example whether monsoon rainfall arises from realistic moisture transport. Treating every model equally, 'model democracy', is simple but ignores differences in quality and independence. Weighting schemes give more weight to models that perform well on relevant metrics and less to near-duplicates. Emergent constraints use a relationship, across models, between an observable quantity and a future response to narrow the projection using observations, but they must have a sound physical basis to be trusted. CMIP6 posed a particular problem: several models had equilibrium climate sensitivity above 5 °C, beyond the AR6 very likely range of 2–5 °C, and these 'hot models' tend to simulate historical warming poorly. AR6 therefore did not take the raw CMIP6 spread as its projection. For global temperature it combined CMIP6 results with observational constraints on past warming and with the assessed climate sensitivity, producing 'assessed' ranges, and it presented regional changes by global warming level (1.5, 2, 3 and 4 °C) rather than only by scenario and year, which reduces the influence of models' differing sensitivities. Users downloading raw CMIP6 data must make similar choices, for example by screening out models whose sensitivity or historical performance is implausible for their application.
Projections need assumptions about future emissions, and these have been organised into scenario families. The IPCC Special Report on Emissions Scenarios (SRES, 2000) described storylines (A1, A2, B1, B2) of population, economy and technology, with emissions derived from them; it was used in the Third and Fourth Assessment Reports. SRES scenarios contained no explicit climate policy. For AR5 the community reversed the sequence to save time. Four Representative Concentration Pathways (RCPs) were defined first by their approximate radiative forcing in 2100: RCP2.6 (a strong mitigation pathway peaking around 3 W m⁻² before 2100 and declining), RCP4.5 and RCP6.0 (stabilisation pathways) and RCP8.5 (a high-emission pathway with forcing still rising). Climate modellers could start running these concentration pathways while social scientists, in parallel, developed the socioeconomic narratives that could lead to them. The number in each RCP name is the forcing in W m⁻² around 2100, not a temperature.
The Shared Socioeconomic Pathways (SSPs), led by O'Neill and colleagues, describe five alternative futures of society without new climate policy, characterised by their challenges to mitigation and to adaptation. SSP1 (Sustainability, 'taking the green road') has low challenges to both. SSP2 (Middle of the road) has intermediate challenges. SSP3 (Regional rivalry, 'a rocky road') has high challenges to both, with slow growth, high population and weak cooperation. SSP4 (Inequality, 'a road divided') has low challenges to mitigation but high challenges to adaptation, because a wealthy elite can decarbonise while much of the population stays vulnerable. SSP5 (Fossil-fuelled development, 'taking the highway') has high challenges to mitigation but low challenges to adaptation, because rapid growth and technology build adaptive capacity while energy demand stays fossil-based. The SSPs are combined with forcing targets in a matrix: a scenario such as SSP2-4.5 means the SSP2 socioeconomic world with climate policy reaching about 4.5 W m⁻² in 2100. AR6 Working Group I used five core scenarios: SSP1-1.9, SSP1-2.6, SSP2-4.5, SSP3-7.0 and SSP5-8.5. Its best estimates of global warming for 2081–2100 relative to 1850–1900 were 1.4 °C for SSP1-1.9 (very likely 1.0–1.8), 1.8 °C for SSP1-2.6 (1.3–2.4), 2.7 °C for SSP2-4.5 (2.1–3.5), 3.6 °C for SSP3-7.0 (2.8–4.6) and 4.4 °C for SSP5-8.5 (3.3–5.7). For the near term (2021–2040), AR6 assessed that 1.5 °C of warming, as a 20-year average, is very likely to be exceeded under SSP5-8.5, likely under SSP2-4.5 and SSP3-7.0, and more likely than not even under SSP1-1.9 and SSP1-2.6. The plausibility of the highest scenarios has been debated. Because of falling renewable energy costs and current policies, many analysts regard SSP5-8.5 and RCP8.5 as unlikely as a 'business as usual' baseline, although they remain useful for exploring high-end physical risk, including strong carbon-cycle feedbacks. Labelling RCP8.5 'business as usual' is therefore misleading. Users should choose scenarios to fit their question: a physical-risk stress test may use a high scenario; a transition-risk analysis should include scenarios consistent with the Paris goals.
Integrated assessment models (IAMs) combine representations of the economy, energy system, land use and, in simplified form, the climate, to explore how emissions might evolve and what mitigation would cost. Two broad types exist. Detailed-process IAMs, such as IMAGE, MESSAGE-GLOBIOM, REMIND-MAgPIE, GCAM, AIM and WITCH, represent technologies, energy supply and demand, and land use in many regions, and compute least-cost pathways to meet a climate target; they produced the SSP baseline and mitigation scenarios and most of the scenarios assessed by IPCC Working Group III. Cost–benefit IAMs, such as DICE, FUND and PAGE, represent damages from climate change as well as mitigation costs in highly aggregated form and are used to estimate the social cost of carbon, the present value of damages from emitting one more tonne of carbon dioxide. Detailed-process IAMs typically assume cost-minimising behaviour under perfect foresight or myopic expectations, and their results depend on assumptions about technology costs, discount rates, the availability of carbon dioxide removal, demand growth and behaviour. They are tools for exploring consistency and trade-offs, not forecasts. Simple climate models or emulators, such as MAGICC and FaIR, translate IAM emissions into concentrations, forcing and temperature, calibrated to match the assessed climate response.
AR6 Working Group III assessed more than a thousand scenarios and grouped them by warming outcome into categories C1 to C8. C1 pathways limit warming to 1.5 °C (with a greater than 50 per cent chance) with no or limited overshoot; C2 return warming to 1.5 °C after a high overshoot; C3 limit warming to 2 °C with a greater than 67 per cent chance; and higher categories correspond to more warming, up to C8 above 4 °C. In C1 pathways, global greenhouse gas emissions fall by about 43 per cent and carbon dioxide emissions by about 48 per cent from 2019 levels by 2030, and net zero carbon dioxide is reached in the early 2050s. In C3 pathways, greenhouse gas emissions fall about 27 per cent by 2030 and net zero carbon dioxide is reached in the early 2070s. All pathways that limit warming to 1.5 °C or 2 °C involve rapid and deep emission cuts in all sectors and some carbon dioxide removal, through afforestation, bioenergy with carbon capture and storage (BECCS) or direct air capture, to counterbalance residual emissions. IAMs have been criticised for leaning on large-scale BECCS, which competes for land and water with food production; for under-representing demand-side measures, behaviour and rapid cost declines in solar power and batteries; for treating damages crudely or not at all; and for using discount rates that make future damages appear small. Newer scenarios include low-energy-demand pathways and more detailed land and food systems. For agriculture, IAM scenarios matter because they imply large changes in land use, bioenergy demand, diets and methane and nitrous oxide emissions, which feed into commodity prices and land competition.
Global models are too coarse for many local decisions: a 150 km grid box can contain mountains, coasts and plains with very different climates. Downscaling derives finer-scale information from coarse model output. There are two main families. Dynamical downscaling runs a regional climate model over the area of interest, driven at its boundaries by GCM output. It produces physically consistent fields of many variables at once and can represent processes, such as orographic rainfall, that statistical methods may miss, but it is computationally expensive, inherits the driving model's large-scale errors, and adds its own biases, so its output usually still needs bias correction. Statistical (empirical) downscaling builds relationships between large-scale predictors and local predictands from observations, then applies them to model output. Methods include regression-based perfect-prognosis approaches, which relate local rainfall or temperature to large-scale circulation and humidity; weather typing and analogue methods, which select past days with similar large-scale patterns; stochastic weather generators, which produce synthetic daily series with statistics adjusted to the projected change; and change-factor (delta) and bias-correction–spatial-disaggregation approaches, which combine interpolation of model change with observed fine-scale climatology. Statistical methods are cheap and can produce large ensembles, but they assume that the observed relationship between predictors and local climate stays the same in a changed climate (the stationarity assumption), and they depend on long, good-quality local observations.
Downscaling adds detail, not necessarily accuracy. The finer output is only as good as the driving model's large-scale circulation and the method's assumptions, and high resolution can give a false sense of precision. 'Added value' should be demonstrated, for example by showing that the downscaled product reproduces observed local climate and variability better than the raw model. Practical choices depend on the question. A crop-impact study needing daily rainfall and temperature at a few sites may use a weather generator or bias-corrected statistical downscaling; a study of monsoon rainfall over mountains may need an RCM. For many applications, ready-made downscaled datasets are available: CORDEX regional simulations; NASA's NEX-GDDP-CMIP6, which provides statistically downscaled and bias-corrected daily data for many CMIP6 models at about 0.25°; and national downscaled products. Users should check which models, scenarios and methods were used, whether the method preserves the model's projected trends, and whether variables are physically consistent with each other. A sound workflow uses several driving models (to sample model uncertainty), several scenarios, and where possible more than one downscaling method, then reports the range. Using a single downscaled model run is one of the most common errors in applied impact studies, because it hides most of the uncertainty that matters for decisions. Pseudo-global-warming experiments, which add a model-derived climate change signal to the boundary conditions of a past event, are a related technique for exploring how a known event would play out in a warmer climate.
Climate model output contains systematic errors, or biases, relative to observations: a model may be 2 °C too warm over a region, simulate too many light-rain days (the 'drizzle effect') or place the monsoon rain belt in the wrong location. Impact models such as crop and hydrological models are sensitive to these biases, so model output is usually adjusted before use. All bias-correction methods compare model output with observations over a common reference period, typically 30 years, and apply the derived adjustment to the future period. The simplest methods adjust the mean. Linear scaling for temperature adds a constant: corrected value = model value + (observed mean − model mean in the reference period). For precipitation it multiplies: corrected value = model value × (observed mean ÷ model mean), which avoids negative rainfall. The delta-change method works the other way round: it takes the observed series and adds the model's projected change in temperature, or multiplies by its projected relative change in rainfall. Delta change preserves the observed day-to-day variability and sequence, which may be an advantage for realism but means any projected change in variability or extremes is lost. Variance scaling corrects the standard deviation as well as the mean.
Quantile mapping corrects the whole distribution. For each model value, its quantile in the model's reference-period distribution is found, and it is replaced by the observed value at the same quantile. This corrects the mean, variance and shape, including the frequency of wet days if a wet-day threshold is applied. A known side effect is that standard quantile mapping can alter the model's projected trend, because the correction applied to future values depends on where they fall in the reference distribution. Trend-preserving variants, such as quantile delta mapping and the ISIMIP3 bias-adjustment method, apply the correction to the change at each quantile and so preserve the model's projected relative changes. Bias correction has limits that users must respect. It cannot correct errors in the model's physics or circulation: if a model places a storm track or monsoon trough in the wrong place, adjusting statistics at a point does not make the projected change trustworthy. It assumes that biases are stationary, the same in the future as in the reference period. Univariate methods correct each variable separately and can break physical relationships, for example between temperature and rainfall on the same day; multivariate methods try to preserve these. Corrections can inflate extremes when applied to coarse grids and then interpolated. And the choice of observational dataset itself matters, because gridded observations disagree, especially for rainfall in data-sparse mountainous regions. Good practice is to evaluate a correction on an independent period, document the method and reference dataset, avoid correcting models whose large-scale behaviour is poor, and treat the corrected output as one of several plausible inputs, not as truth.
Projections are uncertain for three main reasons. Scenario uncertainty arises because future emissions depend on human choices; it is not reducible by better science. Model (or response) uncertainty arises because models differ in structure and parameters and so respond differently to the same forcing, chiefly through feedbacks such as clouds; it can be reduced by better understanding and observations. Internal variability arises because the climate system fluctuates chaotically even with constant forcing, through phenomena such as ENSO and decadal ocean variability; it is irreducible for any particular future year or decade, though it can be quantified. Hawkins and Sutton (2009) showed how the relative importance of these sources changes with lead time and spatial scale. For global mean temperature, internal variability matters most in the next decade or two, model uncertainty is important throughout, and scenario uncertainty grows to dominate by the end of the century. For regional and local quantities, and especially for precipitation, internal variability and model uncertainty remain large for much longer, and in many regions models disagree on even the sign of rainfall change. The smaller the area and the shorter the period, the larger the role of internal variability.
Different ensembles sample different uncertainties. Multi-model ensembles such as CMIP sample model uncertainty. Perturbed-physics ensembles vary parameters within one model. Single-model initial-condition large ensembles (SMILEs), in which one model is run many times with tiny differences in starting conditions, sample internal variability and show how different the next few decades could be by chance alone; they have revealed that observed regional trends over a few decades can differ greatly from the forced response. The time of emergence is the time at which a forced climate signal becomes larger than the noise of internal variability; it comes earlier for temperature than for rainfall and earlier in the tropics, where variability is small, than in the mid-latitudes. Signal-to-noise ratios help decide whether a projected change is robust enough to act on. Where uncertainty cannot be well quantified, other approaches help. Storylines describe physically self-consistent, plausible futures, for example a drying and a wetting storyline for a region, to explore consequences without assigning probabilities. Robust decision making and adaptation pathways, described in the adaptation subject, test decisions across many plausible futures. IPCC calibrated language (likely, very likely; low, medium, high confidence) should accompany any statement drawn from the assessments. The cardinal errors in practice are to present one model run as 'the future', to quote a mean without its spread, and to treat projected change for a single year or grid cell as a forecast.
Weather forecasts are an initial-value problem: they start from the observed state of the atmosphere and integrate forward. Because the atmosphere is chaotic, as Edward Lorenz showed in the 1960s, small errors grow, and deterministic skill for day-to-day weather is lost after roughly two weeks. Beyond that, useful predictability comes from slowly varying parts of the system that influence the atmosphere: sea surface temperatures, especially ENSO in the tropical Pacific and the Indian Ocean Dipole; soil moisture; snow cover; sea ice; and the stratosphere. Seasonal forecasts exploit these boundary influences to predict shifts in the probabilities of seasonal averages, such as a wetter or drier than normal season. Sub-seasonal to seasonal (S2S) forecasts, with lead times of about two weeks to two months, bridge weather and seasonal ranges. A key source of skill at this range is the Madden–Julian Oscillation (MJO), an eastward-moving pattern of tropical convection with a period of roughly 30–60 days that modulates monsoon active and break spells and tropical cyclone activity. Extended-range forecasts help with decisions such as sowing, irrigation scheduling and harvest planning. ENSO is monitored with sea surface temperature indices; the Niño 3.4 region spans 5°N–5°S and 170°W–120°W, and the United States NOAA's Oceanic Niño Index uses a three-month running mean of its anomaly, with El Niño conditions defined at +0.5 °C or more and La Niña at −0.5 °C or less, sustained over several overlapping seasons. Other agencies use different thresholds, so ENSO statements should name their source.
Seasonal forecasts are produced with coupled ocean–atmosphere models run as ensembles from slightly different initial conditions, and with statistical models that relate predictors such as sea surface temperatures to seasonal rainfall. Multi-model ensembles, such as those combined by the WMO Lead Centre for Long-Range Forecast Multi-Model Ensemble and the Copernicus Climate Change Service, are generally more reliable than single models. Forecasts are usually expressed as probabilities for tercile categories: below normal, near normal and above normal, each with a climatological probability of one third. Regional Climate Outlook Forums, such as those for South Asia and the Greater Horn of Africa, bring together national meteorological services to produce consensus outlooks. As a national example, the India Meteorological Department issues long-range forecasts of southwest monsoon rainfall in stages, in April and in late May or June, using a monsoon mission coupled model and statistical models. Seasonal rainfall is expressed as a percentage of the long period average (LPA), which is about 87 cm for 1971–2020; 96–104 per cent of LPA is classed as normal, below 90 per cent as deficient, 90–95 per cent below normal, 105–110 per cent above normal, and above 110 per cent excess. Forecasts must be verified. The Brier score for a probability forecast of a binary event is the mean of (p − o)² over all forecasts, where p is the forecast probability and o is 1 if the event happened and 0 otherwise; lower is better, and 0 is perfect. The Brier skill score compares it with a reference such as climatology: BSS = 1 − BS/BS_ref, where positive values show skill. Reliability diagrams test whether events forecast with, say, 70 per cent probability happen about 70 per cent of the time; the relative operating characteristic (ROC) measures the ability to discriminate events from non-events. Seasonal skill is highest in the tropics, especially during strong ENSO events, and lower in the mid-latitudes.
Raw daily data are summarised into indices that describe aspects of climate relevant to impacts. The Expert Team on Climate Change Detection and Indices (ETCCDI), set up by WMO's Commission for Climatology, CLIVAR and JCOMM, defined 27 core indices of temperature and precipitation extremes that are widely used for monitoring and in IPCC assessments. Temperature indices include TXx (annual maximum of daily maximum temperature), TNn (annual minimum of daily minimum temperature), SU (summer days, daily maximum above 25 °C), TR (tropical nights, daily minimum above 20 °C), FD (frost days, daily minimum below 0 °C), ID (ice days), GSL (growing season length), the percentile-based TX90p (share of days with maximum above the 90th percentile of a base period) and TN10p, the warm and cold spell duration indices WSDI and CSDI, and the diurnal temperature range DTR. Precipitation indices include Rx1day and Rx5day (maximum one-day and five-day totals), SDII (simple daily intensity index, the total rainfall on wet days divided by the number of wet days, with a wet day defined as at least 1 mm), R10mm and R20mm (counts of days with at least 10 mm and 20 mm), CDD (maximum number of consecutive dry days, with less than 1 mm), CWD (consecutive wet days), R95p (rainfall on very wet days above the 95th percentile) and PRCPTOT (total wet-day rainfall). Fixed-threshold indices, such as SU and FD, are easy to interpret but suit some climates poorly (a 25 °C threshold is exceeded on almost every day in the tropics); percentile-based indices are comparable across climates but need a base period and careful handling at its edges. Many sector-specific indices exist, such as the Expert Team on Sector-specific Climate Indices (ET-SCI) set for health, agriculture and water.
Drought indices such as the SPI and SPEI, covered in the adaptation and impacts subjects, express precipitation or water balance anomalies in standardised form. Agricultural indices include growing degree days, frost-free season length, crop water stress and the start and end of the rainy season defined by rainfall accumulation rules. Heat-stress indices combine temperature with humidity, and sometimes radiation and wind: the heat index, the wet-bulb globe temperature (WBGT), used in occupational standards, and the Universal Thermal Climate Index (UTCI). For livestock, the temperature–humidity index is used. Large-scale modes of variability are tracked with indices that feed forecasts and risk analysis: Niño 3.4 and the Oceanic Niño Index for ENSO; the Southern Oscillation Index, based on the pressure difference between Tahiti and Darwin; the Dipole Mode Index for the Indian Ocean Dipole, the difference in sea surface temperature anomaly between the western and south-eastern tropical Indian Ocean; and the North Atlantic Oscillation index. When computing indices from model output, users should compute them from daily data for each model and then aggregate, rather than computing from an ensemble-mean series, because averaging smooths out extremes. Indices from bias-corrected and raw data can differ substantially, particularly for threshold-based rainfall indices affected by the drizzle effect. Comparisons between observed and modelled indices must use the same definitions, base periods and spatial scale, because station extremes are larger than grid-box averages.
Practitioners need three kinds of data: observations, reanalyses and model projections. Station observations come from national meteorological services and are compiled into global datasets, such as those of NOAA's National Centers for Environmental Information. Gridded observational products interpolate stations onto regular grids: the Climatic Research Unit's CRU TS monthly dataset; the Global Precipitation Climatology Centre (GPCC) rainfall analyses; national grids such as IMD's 0.25° daily rainfall for India; and satellite–station blends such as CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data), which covers 50°S–50°N at 0.05° resolution from 1981 and is widely used for drought monitoring in Africa and Asia. WorldClim provides high-resolution long-term climatologies. A reanalysis combines a fixed weather-forecast model with historical observations through data assimilation to produce a complete, consistent gridded record of the past atmosphere. ERA5, from the European Centre for Medium-Range Weather Forecasts under the Copernicus Climate Change Service, provides hourly data at about 0.25° (31 km) from 1940 to near the present; ERA5-Land offers finer land variables. Reanalyses are invaluable where stations are sparse, but variables poorly constrained by observations, such as rainfall, carry model biases. CMIP model output is distributed through the Earth System Grid Federation (ESGF) and, for many variables, through the Copernicus Climate Data Store. Downscaled products include CORDEX and NASA's NEX-GDDP-CMIP6. Agricultural users often use NASA POWER, which provides daily meteorological and solar data formatted for crop models.
Several portals make data and summaries accessible to non-specialists. The IPCC WGI Interactive Atlas lets users map observed and projected changes for many variables and regions, by scenario or global warming level, with the underlying data available. The World Bank Climate Change Knowledge Portal provides country and subnational summaries. The Copernicus Climate Data Store hosts reanalyses, seasonal forecasts, projections and sectoral indicators with a toolbox for processing. National portals, such as those of national meteorological agencies, give downscaled projections tailored to national needs. Good practice when using these resources includes: reading the documentation to understand the variable definitions, units, calendars (some models use 360-day or no-leap calendars), grids and reference periods; checking licences and citing datasets with their version and digital object identifier; choosing an ensemble of models and scenarios rather than one; comparing a model's historical simulation with observations for your region before trusting its projections; keeping a record of processing steps so analyses are reproducible; and following FAIR data principles (findable, accessible, interoperable, reusable). Data formats such as NetCDF with CF (Climate and Forecast) conventions, and tools such as CDO, NCO, xarray in Python and the terra and ncdf4 packages in R, are standard. Be careful about combining datasets with different resolutions or definitions, about using station data with inhomogeneities from instrument or site changes, and about comparing grid-box averages with point observations. When an answer depends strongly on the dataset chosen, report the results from several datasets so that users can see the structural uncertainty.
Companies, banks, insurers and public bodies use climate scenarios to test how their strategies, assets and portfolios would fare under different plausible futures. The Task Force on Climate-related Financial Disclosures (TCFD), whose recommendations were published in 2017, organised disclosure around governance, strategy, risk management, and metrics and targets, and asked organisations to describe the resilience of their strategy under different climate-related scenarios, including a 2 °C or lower scenario. The International Sustainability Standards Board's IFRS S2 Climate-related Disclosures, issued in 2023, builds on TCFD and requires entities to use climate-related scenario analysis to assess climate resilience, proportionate to their circumstances. Central banks and supervisors run climate stress tests of the financial system. Climate-related risks are usually divided into physical and transition risks. Physical risks are acute, from extreme events such as floods, cyclones and heatwaves, or chronic, from longer-term shifts such as rising temperatures, sea-level rise and changing rainfall. Transition risks arise from the move to a low-carbon economy: policy and legal (carbon pricing, regulation, litigation), technology (substitution by low-carbon alternatives), market (shifts in demand and prices) and reputational. The two are linked inversely in scenarios: strong mitigation raises transition risk and lowers long-run physical risk, while weak mitigation does the opposite. Opportunities, such as resource efficiency and new markets, are also part of the analysis. Scenarios are not forecasts or predictions. They are plausible, internally consistent descriptions of how the future might unfold, chosen to span relevant uncertainties. Exploratory scenarios ask 'what if?'; normative scenarios start from a goal, such as net zero by 2050, and work backwards.
Most organisations start from published reference scenarios rather than building their own. The Network for Greening the Financial System (NGFS), a group of central banks and supervisors, publishes scenarios built with IAMs and grouped into categories: orderly (such as Net Zero 2050 and Below 2 °C), in which climate policies are introduced early and become gradually more stringent; disorderly (such as Delayed Transition), in which policies are delayed or divergent and transition costs are higher; and hot house world (such as Current Policies and Nationally Determined Contributions), in which limited action leads to high physical risk. Later vintages add a 'too little, too late' category. The International Energy Agency's scenarios, including Net Zero Emissions by 2050, Announced Pledges and Stated Policies, are widely used for energy-sector transition analysis. For physical risk, organisations usually draw on SSP–RCP climate projections, often a low and a high scenario such as SSP1-2.6 and SSP5-8.5, at time horizons such as 2030, 2050 and 2100. A typical process is: identify material climate risks and opportunities; choose scenarios and time horizons covering a range of outcomes; translate scenario variables, such as carbon prices, commodity demand, temperature or flood frequency, into impacts on revenues, costs, asset values and supply chains; assess resilience and identify responses; and disclose assumptions and results. For an agri-food company, a physical scenario might examine heat and drought effects on sourcing regions using crop-model results, while a transition scenario might examine carbon pricing of methane and fertiliser emissions, deforestation regulation and changes in diets. Common pitfalls are treating one scenario as the expected future, using average changes that hide extremes, ignoring compound and cascading risks, picking time horizons too short to show physical risk, and presenting precise financial numbers from highly uncertain chains of models. The value of scenario analysis is in the strategic conversation it forces, not in a single number.
A student uses this subject to understand where the numbers in impact papers come from and to pass questions that test the scenario frameworks (what the number in RCP4.5 means, how SSP4 differs from SSP5), the model hierarchy, the meaning of downscaling and bias correction, the three sources of uncertainty, ETCCDI definitions and forecast verification. The numerical core is the energy-balance response time, simple bias corrections, extreme indices from a daily series, and Brier scores. A research student also learns to download CMIP6 or CORDEX data, bias-correct it against a gridded observation set and run an ensemble through an impact model.
A five-year practitioner, such as an agrometeorologist, a climate-risk analyst at a bank or an adaptation consultant, uses the subject to choose models, scenarios and downscaled datasets for a project; to bias-correct and quality-check data; to compute indices relevant to the client, such as heat days for a dairy or consecutive dry days for a rainfed district; to explain seasonal outlooks probabilistically; and to build physical and transition scenarios for TCFD- or ISSB-aligned disclosure. The key judgement is how much uncertainty to carry through and how to communicate it.
A senior professional with 25 years' experience, such as a chief risk officer, a head of climate services at a meteorological agency or a ministry adviser, uses the subject to commission scenario analyses and climate services, to challenge consultants who present a single model run or a high-end scenario as 'business as usual', to choose scenario sets that span both physical and transition risk, and to ensure that board decisions reflect uncertainty honestly. Such a person also sets data and method standards across an organisation.
A researcher with a PhD and 25 years' experience works on improving and evaluating models, developing downscaling and bias-correction methods that preserve physical consistency, designing large ensembles and emergent constraints, advancing seasonal and sub-seasonal prediction for agriculture, and building scenario frameworks that link climate, land use and food systems. Such a researcher contributes to CMIP, CORDEX, IPCC assessments and national climate projections, and mentors teams in reproducible climate-data science.
This manual does not teach the equations of GCMs, numerical methods or data-assimilation algorithms, and it names models and datasets without ranking them. Scenario vintages (NGFS, IEA) and disclosure rules (ISSB, regional regulations) change frequently; their current categories and requirements should be checked. CMIP7 experimental design, the AR7 scenario framework and the latest seasonal forecast systems are described only in general terms because they were still evolving. It gives no country-by-country projections and no guidance on the financial modelling that converts scenario variables into asset values, which belongs to the climate-finance subject. Specific skill scores of seasonal forecasts are omitted because they vary by region, season and system.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.