Migration Anticipation and Preparedness

4. Forecasters’ toolkit: Choosing the right model for the task

What are the existing forecasting methods and models?

There are various approaches to migration forecasting, both quantitative and qualitative, each with distinct methodological assumptions, strengths and shortcomings. This chapter presents an overview of key forecasting methods, which serve as the methodological foundation for models featured throughout this handbook. For comprehensive overviews, see Bijak (2011[1]), Sardoschau (2020[2]) Wiśniowski (2021[3]), Disney et al., (2015[4]), Bijak et al (2019[5])). For a detailed description of the terms used in this publication, please refer to the glossary.

Judgemental scenarios or deterministic models are commonly used in demographic forecasting to outline plausible future trajectories for population components – fertility, mortality, and migration. They can also be used, for example to anticipate future family migration based on simple chain migration models. These approaches are typically grounded in past demographic trends and can be informed further by qualitative insights in case of major institutional changes (e.g. 2004 EU enlargement). They produce baseline, high and low trajectories, but these are not probabilistic forecasts (see Box 4.2 for a recent example of demographic analysis based on probabilistic projections of migration). These methods generally do not account for unforeseen changes in government policy, economic shocks, or environmental disruptions, and they do not explicitly model uncertainty.

Expert-based forecasts incorporate qualitative insights on future developments – migration policies, economic trends, environmental changes, and technological advances – that may not be reflected in historical data (see for example Lutz and Scherbov (1998[6]) or Sohst, Acostamadiedo and Tjaden (2023[7]) see also Box 4.1). Expert judgement may be provided directly by the forecasters or elicited through structured expert surveys such as the Delphi method (see Chapter 6 and Box 6.2). The Delphi method facilitates iterative feedback and consensus building in an anonymous setting. While useful, such methods can suffer from overconfidence and a tendency to understate uncertainty – especially when trends have been relatively stable.

Box 4.1. Expert-based asylum forecasts: The example of Switzerland

Switzerland’s (State Secretariat for Migration) has had a migration analysis unit specialising in asylum applications forecasts since 2005. Since asylum migration can change almost overnight by events that are nearly impossible to predict (COVID‑19, Araspring, etc.), Switzerland chose to establish migration flow scenarios based on expert judgment without using data models. Scenarios are built from six broad factors:

1. The situation in countries of origin (e.g. conflict, human rights violation, economic pressure to migrate) is collected by migration analysts through a survey. Country-of-origin experts from the SEM are asked to evaluate the likely evolution of flows to Europe. Further reports of Migration liaison officers stationed in Swiss embassies are taken into consideration;

2. The situation on migration routes (e.g. quantitative information on irregular border crossings, migration control measures taken at the EU external border, legal travel options) are gathered from countries and other international stakeholders;

3. The economic situation in EU countries of first entry vis-à-vis Switzerland;

4. Measures taken by European countries in terms of migration and asylum policy;

5. Measures taken by Switzerland in terms of migration and asylum policy;

6. Future Swiss implementation of the Dublin procedure and the European Pact on Migration and Asylum.

These six factors are given a different “weight” e by migration analysts. At the beginning of each year, based on all these information inputs, the migration analysis unit estimates four scenarios. The experts assign different probabilities of occurrence, as well as the main nationalities of origin of applicants. This yearly prognosis and its scenarios are then re‑evaluated by mid-year.

Time series models, particularly autoregressive integrated moving average (ARIMA) models, extrapolate past migration trends into the future. These models assume that future values depend on past values, a drift term, and an error component (see for example de Beer, J. (1997[8]) or Keilman et al. (2001[9])). For example, a random walk with drift (ARIMA(0,1,0)) is often used to model migration trends: ln(m_t+1) = c + ln(m_t) + ε_t, where the logarithm of predicted migration, m_t+1, depends on the value from the preceding period, m_t, a drift constant c, and an error term, ε_t, usually assumed to follow independent normal centred distributions. Multivariate extensions, such as ARIMAX or Vector Auto-Regressive (VAR) models allow inclusion of additional interdependent variables such as GDP growth or unemployment rates. Time series models are attractive due to their well-established statistical foundations and available software tools. However, they are mostly extrapolative and cannot easily incorporate policy or structural changes unless pre‑specified.

Expert-based probabilistic population projections have the form where the migration flow under study, $v_t$ is composed of the average trajectory of the process, $\stackrel{-}{v}$, which is an a priori assumption taken from the subjective expert judgment and the error term, ${\epsilon }_t$, following the chosen stochastic process (see for example Lutz et al. (2000[10])). Using expert elicited opinions, stochastic processes are then run to generate a range of possible future population scenarios. Each simulation accounts for random variations in demographic rates, reflecting the inherent uncertainty in population dynamics. These forecasts provide a full probability distribution of future population sizes and structures, rather than a single deterministic prediction. Such models are especially limited by their lack of empirical grounding, though adjustments may be made when structural breaks (e.g. EU enlargement) are anticipated.

Agent based models simulate individual behaviours and decisions within a system of interacting agents – migrants, institutions, or other actors. Agents respond to information (e.g. wage differentials) and make decisions stochastically in an imperfect information environment (see for example Willekens (2019[11]) and Klabunde and Willekens (2016[12]) or McAlpine et al. (2020[13])). These models allow for rich causal modelling of migration dynamics, life‑course trajectories, and interactions between micro- and macro-level factors. ABMs can incorporate feedback between individual decisions and institutional responses, offering a powerful lens for causal and scenario-based forecasting. The main drawback of these models is that they rely on extensive hypotheses and parameters that are hard to estimate and have a large impact on outcomes. They are usually more useful to identify transmission channels to dependent variables in complex interactive systems rather than produce reliable forecast.

Econometric or causal models, such as logit or probit models, estimate migration as a function of explanatory variables, enabling theory testing and policy analysis. Migration theories are however at this stage too fragmented and unsettled to support forecasting.

Gravity Models are among the most widely used causal type of models (see Beine et al. (2015[14]) for a comprehensive overview). They are based on distance, previous migration waves as well as origin destination macroeconomic characteristics and linkages. They explain past migration flows with a satisfactory degree of accuracy but usually fail to anticipate breaks in migration trends and are a poor fit for forced displacement. Multilateral resistance in gravity models makes it possible account for alternative destinations, which may be particularly relevant for some categories of migration where competition between destination countries is strong (e.g. international students or highly skilled migration). Recent work (Beyer, Schewe and Lotze‑Campen (2022[15]), Welch and Raftery (2022[16])), however, show that gravity models can neither explain nor correctly predict the temporal dynamics of human movement.

Another example of econometric models used in migration forecasting is Simultaneous Equations Models (SEMs) which model multiple interdependent relationships between migration and other socio‑economic factors, offering richer system-level analysis (see for example Burzynski et al. (2020[17]); Dao et al. (2021[18])). While promising for regional forecasting, SEMs are constrained by data comparability and availability across countries.

Bayesian approach incorporate prior information – including expert judgement – into statistical models (see for example model (Bijak and Wiśniowski (2010[19])). Forecasting in the Bayesian approach is based on the construction of a probability distribution of the vector of future values of the dependant variable, X_f, conditional on the vector of past (observed) values, X, and taking into account the posterior knowledge on the parameters of the forecasting model, θ. The predictive probability distribution of X_p can be interpreted as an average from the conditional predictive distribution p(X_p|θ, X), weighted with the posterior probabilities of the parameters (see Bijak (2005[20]) for more detailed explanation). An important issue is the selection of the prior probability distribution of the estimated parameters, p(θ), reflecting the knowledge of the researcher. This approach allows coherent integration of model uncertainty, data uncertainty and inherent uncertainty about future conditions. Bayesian methods are especially valuable when historical data are sparse or noisy. For example, if an expert believes a given flow is non-stationary, the prior distribution for an autoregressive term in the model takes a value of 1 or above. See Chapter 6 for more details).

Machine learning models are data-driven rather than assumption-driven and are increasingly used for short-term (days to months) migration forecasting (see for example Carammia et al. (2022[21]); Robinson and Dilkina (2018[22])) For example, Dynamic Elastic Net (DynENet) is a relatively new type of regularisation method that combines model selection and estimation within a single framework. It offers: parsimony (selects only relevant variables), flow-specific adaptability, resilience to missing data, time adaptability (via moving training windows) and interpretability through its linear functional form (Carammia, Iacus and Wilkin, 2022[21]).

Another example of machine learning models is Long Short-Term Memory (LSTM) Networks, which is a form of recurrent neural networks (RNN) capable of learning long-term dependencies in sequential data. They maintain and update “memory” states to inform predictions. Though powerful in capturing non-linearity, LSTMs are often criticised for limited interpretability and a tendency to overfit.

Box 4.2. Probabilistic projections of future trends in international migration in World Population Prospects 2024

For the first time, the medium variant of net international migration in the 2024 revision of the United Nations World Population Prospects was projected using probabilistic methods. A Bayesian hierarchical approach was used to model future trajectories of the crude net migration rate and to account for uncertainties and correlations across countries and over time. The analytical approach follows these 10 steps:

1. Using the annual estimates of crude net migration rates from 1950 to 2023 for all countries from the 2024 revision, a Bayesian hierarchical first-order autoregressive model, or AR(1), was fitted to crude net migration rate data for all countries to estimate the parameters of the Azose and Raftery (2015[23]) migration model.

2. By default, the estimation period used for model fitting referred to 1950 to 2023, but the application of autoregressive statistical methods in this context faced two types of challenges for some countries: (1) the country-specific overall rate of change and long-term mean net migration rate are influenced by the reference period used to fit the model, which might need to be adjusted depending on the country’s historical context (e.g. conflicts or political and economic changes), and (2) the short-term and long-term projections are highly influenced by the most recent situation (e.g. countries experiencing large in/out flows of migrants or refugees in 2023). Expert knowledge was incorporated to inform prior distributions and provide further guidance to the model fitting with respect to the reference period used for a specific set of locations.

3. To address the challenges with the set of countries experiencing recent large movements of net international migrants or refugees, expert knowledge was used to define the expected target years when the situation in these locations is expected to return to some baseline level. For a subset of those that experienced large in/outflow of recent refugees, expertise was also used to determine what proportions of refugees are expected to return to their country of origin by the target year. For each of these cases, some linearly interpolated values were estimated between the 2023 baseline and the expected median value for the user-defined target year. All probabilistic trajectories were adjusted proportionately based on the expected values for the median trajectory of each country.

4. Using the results from the model fitting (step 1), and the additional constraints (steps 2 and 3), the R package bayesMig (Azose, Sevcikova and Raftery, 2023[24]) was used to create projections of net migration rates, and generate 2 000 trajectories from 2024 to 2100 for each country.

5. Upon review of the median projection results for the net migration rates for each country, for a subset of locations, the default level for the long-run projection up to 2100 was considered to deviate too much from prior expectations and expert knowledge about country history and context. For each of these cases, user-defined target values were defined and used as constraints for the user-defined projection period. All probabilistic trajectories were adjusted based on these expected values for the median trajectory of each country.

6. Additional constraints were also imposed on the cumulative migration rates over different time periods to prevent net migration from growing indefinitely, which would generate unrealistic population growth. The following upper-bound constraints were imposed on cumulative migration for each of the 2 000 trajectories of several countries of the Gulf region, where temporary labour migrants comprise a large share of the population. These limits were set based on past fluctuations experienced by these countries and the unconstrained 95th percentiles.

7. Based on the additional constraints from steps 5 and 6, a new set of projections of net migration counts was generated to consider these set of constraints in addition to those from steps to 2‑3, and 2 000 trajectories were exported for each country.

8. Using these 2 000 trajectories of net migration rates, an initial set of net migration counts was computed to generate probabilistic population projections based on the population, fertility and mortality data by age and sex from the 2024 revision, and the 2 000 trajectories of total fertility rates, life expectancy at birth by sex, and net migration rates (from step 6) combined with model-based assumptions about the expected distribution of migrants by sex and age.

9. As with estimates, net migration balancing was carried out for projections as a final step to ensure that, at the global level, the sum of all net international migration flows equalled zero for each year of the projection period. To ensure that for each of the 2 000 trajectories the sum annually of net migration counts across all countries is zero, an extra step was implemented to “balance” total net migration counts.

10. The 2 000 adjusted and balanced trajectories of total net migration counts (from step 9) for each country were exported and used as inputs to the final run of the probabilistic population projections. They were also used to generate the median projected values for each country that were used for the medium deterministic scenario of the 2024 revision (after applying an additional rebalancing step as described in step 9 because the annual sum of all the country medians is initially unbalanced).

Source: UN DESA/POP/2024/DC/NO.10 (2024) Methodology of the United Nations population estimates and projections, https://population.un.org/wpp/assets/Files/WPP2024_Methodology.pdf.

What is the appropriate forecasting model for each migration category?

The previous sections provide an overview of the different forecasting methods. Various criteria – such as the type of migration flow, the stationarity of the data series, the time horizon, and available resources – help determine the most appropriate forecasting model among them (see Figure 4.1). For example, for forced migration (irregular border crossings, asylum, as well as categories in-between regular and irregular, such as mixed migration phenomena), machine learning models using both traditional and innovative data are preferred, whereas for regular migration, standalone econometric models or econometric models with expert opinions can be utilised.

Figure 4.1. Decision tree to identify the best migration forecasting models according to national practitioners’ needs and resources

Where previously the dominant strand of the literature was on overall migration trends, in the last decade, attention has shifted towards forced displacement and irregular border crossings. In the meantime, the deterministic models and expert forecast as well as time series and causal models have been progressively supplemented by machine learning and Bayesian models that have become more operational as big data and large computational power became more accessible. This is true for short term forecasts such as nowcasting (Box 4.3) as much as for medium (12‑18 months) forecast.

Box 4.3. Nowcasting: Short-time horizon forecasts and early warning systems

A special category of migration forecasting methods concerns approaches to nowcasting – predicting migration flows for the current period or the nearest future. This strand of research has seen rapid developments in the recent years, amplified by the proliferation of large‑scale data sources, including mobile phone data (Lai et al. (2019[25])), social media (Alexander et al. (2019[26]); Culora et al. (2021[27]); Leasure et al. (2023[28]); Rampazzo et al. (2024[29])), online search data (Böhme et al. (2020[30]); Avramescu and Wiśniowski (2021[31])) or experimental prediction markets (Morgenstern and Strijbis (2024[32])). Many of these methods have successfully attempted combining the advantages of the non-traditional sources, such as the timeliness of mobile phone or social media data, with the known features of more traditional data, such as censuses or surveys. This area of application of novel methods to predicting current or near-future migration flows is very promising, and has already gained traction also within the user community: organisations such as the International Displacement Monitoring Centre (IDMC (2022[33])) or the UN Refugee Agency (UNHCR (2025[34])) are developing their own nowcasting models for the populations of interest – particularly important for such rapidly-changing contexts as asylum or other forced displacement categories.

A special group of related methods, closely linked with the notion of preparedness for rare events (for discussion, see Bijak (2024[35])), concerns early warnings – detection of changes in the trends in migration or displacement flows, typically based on higher-frequency data (daily, weekly or monthly), such as asylum-related flows. Such techniques are aimed at identifying the potential changes, ideally contemporaneously or ahead of time, to flag them for further attention and possibly for operational response. The methods used for early warning applications include simple, descriptive techniques, such as cumulative sum (cusum) approaches (Napierała et al., 2021[36]) the momentum approach (Carammia, Iacus and Wilkin, 2022[21]) or the use of alerts when flows exceed a high percentile threshold during the preceding period. They also range from econometric models (Barker and Bijak, 2022[37]), to more advanced machine learning techniques, which at the same time aim to identify the signal indicating a possible change in trends, as well as selecting the most meaningful predictors from a range of options, including large‑scale text databased (such as GDELT – Global Database of Events, Language and Tone, http://www.gdeltproject.org/, see Carammia, Iacus and Wilin (2022[21])). Early warning models are typically evaluated based on how well they are able to predict the actual changes in trends, avoiding both false alarms and missed detections.

The current methodological state of the art of migration forecasting is one of continuous innovation, constrained, however, from one side by the unavailability of adequate and appropriate data, and from the other side by the inherent uncertainty of migration processes. With these two constraints in mind, one fundamental methodological question in anticipating migration flows is, how to choose the best bespoke model for a task at hand? This section briefly summarises different classes of models that might be relevant for different types of migration and time horizons. It also examines the relative advantages and drawbacks of different types of models, taking into account the available resources – financial, human skills, equipment and so on – as well as data availability.

When discussing the choice of the right modelling approaches, it can help to start from the features of the process at hand. From a statistical viewpoint, one such key feature is stationarity – a property indicating whether the migration process has the same characteristics (such as mean or variability) over time (Bijak, 2011[1]; 2024[35])). Stationary processes are easier to model, and their predictive uncertainty is more straightforward to describe by using probabilities, and to calibrate based on past time series data. Stationarity indicates the presence of an equilibrium, to which migration eventually returns even after events generating much higher or lower flows. On the other hand, for non-stationary processes – such as random walks, processes with trends or level shifts – the equilibrium is constantly changing. As a result, the predictive uncertainty is increasing rapidly, rendering forecasts of little use beyond the horizon of a few years ahead (Bijak and Wiśniowski, 2010[19]).

The distinction between stationary and non-stationary processes is visible in the analysis of different types of migration flows, such as irregular border crossings, asylum, family, study, work, or returns. As argued by de Beer (2008[38]), such flows should ideally be analysed and forecast separately, as they are likely to be responding to different drivers and policies and be characterised by different dynamics. Treating them separately not only makes it possible to formally address the heterogeneity of migration flows, but also to find the best-matching statistical model in each case. This suggestion has been confirmed empirically: Bijak et al. (2019[5]) analysed several flows of migration into the United Kingdom in the early 2000s (returns being less volatile, labour migration more volatile, and asylum migration the most), ending up with concrete recommendations for model choice guided by the nature of specific processes and the availability of data. Crucially from the point of view of the practical uses of forecasts, Bijak et al. (2019[5]) suggested adopting a risk management approach, whereby the uncertainty (and unpredictability) of the processes is juxtaposed with their impact for a given stakeholder, to allow prioritising resources in the high-uncertainty and high-impact areas, such as asylum.

The practical recommendations for the choice of an appropriate method for forecasting are therefore flow-specific: Bijak et al. (2019[5]) suggested proceeding from (a) the analysis of statistical properties of the flows of interest (such as stationarity or uncertainty), through (b) the formal assessment of the available data (administrative registration or censuses vs. novel data and digital traces); to (c) selection of an appropriate model that is capable of reflecting the properties (a) and fits to available data (b) (See Figure 4.1).

Table 4.1 summarises the stylised characteristics of seven main types of migration flows, alongside the assessment of their properties and typically used data, and recommendations for method selection, based on the arguments presented above. How to precisely develop the most efficient model is discussed later (see Chapter 6). Unsurprisingly, volatile processes such as irregular border crossings and asylum exhibit non-stationarity and high uncertainty but given the availability of relatively long series of higher-frequency data, offer the possibility of using advanced machine learning methods, unlike most of the other flows, which are measured only quarterly or annually. The potential role of machine learning approaches for regulated migration streams where high-frequency administrative and linked datasets are becoming increasingly available have been barely investigated so far (see Box 6.3 for the use of a machine learning approach to forecast international student flows). It remains to be seen if those approaches may be useful for less volatile and uncertain migration categories.

Table 4.1. Stylised characteristics of seven main types of migration flows, typical data, and recommendations for selecting possible forecasting methods

Migration type	Stationarity	Uncertainty	Typical data sources	Frequency of data	Length of series	Possible methods
Irregular border crossings	Highly unlikely	Very high	Administrative data, novel data & “digital traces”	Typically daily, weekly, or monthly	Long: several years of weekly data	Statistical & machine learning; expert-based
Asylum	Highly unlikely	Very high	Administrative data, novel data & “digital traces”	Typically daily, weekly, or monthly	Long: several years of weekly data	Statistical & machine learning; expert-based
Family	Likely	Medium	Registers, administrative data, surveys	Typically quarterly or annual	Short: up to 10‑20 years	Statistical & econometric time series
Study	Unlikely	Medium	Registers, administrative data, surveys	Typically annual	Short: up to 10‑20 years	Statistical & econometric time series; expert-based
Labour	Unlikely	High	Registers, administrative data, surveys	Typically quarterly or annual	Short: up to 10‑20 years	Statistical & econometric time series; expert-based
Return	Likely	Low to medium	Registers, censuses, surveys	Typically quarterly or annual	Short: up to 10‑20 years	Statistical & econometric time series
Transit	Unlikely	Medium	Registers, surveys	Typically quarterly or annual	Short: up to 10‑20 years	Statistical & econometric time series; expert-based

In OECD countries and international organisations involving OECD countries, approximately half of the forecasting exercises combine statistical/econometric time series models with expert opinions (Table 4.2). These models rely on past data, such as asylum applications or residence permits, alongside non-quantitative sources like policy changes, intelligence reports, and scenario planning (see Table 5.1 for more details on data sources). In contrast, over one‑third of forecasting approaches rely exclusively on data-driven models, while expert-based models are used by only three countries, notably in Switzerland.

For forecasting forced migration, most countries employ the mixed-method approach, while forecasting regulated migration flows is more varied. For the latter, the number of mixed approaches is roughly equal to those based purely on quantitative methods (four countries each), with two countries using expert elicitation only. Notably, very few countries and EU agencies have explored the use of machine learning models for migration forecasting so far.

Table 4.2. What models are currently used in OECD countries?

Institution	Migration type	Model
Frontex	Border Crossings	Methods including time series and regression analysis, historical analogy, panel consensus and qualitative risk management analysis.
Germany	Border Crossings	Methodologies incorporating AI and machine learning technologies (still in its nascent stage) have been used within a 12‑month timeframe.
United States MAC	Border Crossings	Mixed-methods framework that combines machine learning models with judgmental adjustments from subject matter experts (SMEs).
Austria	Asylum Applications	Extrapolations of observed trends using scenario assumptions.
Belgium	Asylum Applications	Time series regression leading to 3 scenarios, with a seasonal effect included. Scenarios are then discussed by experts.
EUAA	Asylum Applications	Several models (e.g. TSLM, exponential smoothing, ARIMA, Machine Learning, time series regression based on month-to-month percentile changes in asylum applications over the last 6 years) are tried and their results are used to form 4 to 5 scenarios, from the (very) lowest to the (very) highest scenario. The likelihood of the scenarios built from the models are decided with expert elicitation.
France	Asylum Applications	DSED/DGEF: Linear, log or exponential regression model based on past asylum applications, adjusted after an expert meeting. OFPRA: Linear or log regression model based on past asylum application changes.
Germany	Asylum Applications	Scenario-based forecast adapted from the EUAA methodology (see above).
Ireland	Asylum Applications	Scenario-based forecast adapted from the EUAA methodology (see above). This approach includes five scenarios, ranging from a “low” scenario (20th percentile changes) to a “high” scenario (80th percentile changes). Each one of these is based on historical data, supplemented by an additional data-driven scenario based on Seasonal and Trend decomposition.
Netherlands	Asylum Applications	Bayesian probabilistic forecasting combination of statistical time series, and harmonised expert intuition.
Norway	Asylum Applications	Regression models based on recent data and predictive factors that may affect future flows.
Poland	Asylum Applications	Method that accounts for the seasonality of migration phenomena based on the analysis of historical data and the knowledge of employees and collaborators. This includes information about countries of origin, diasporas in Poland and other European countries, and current migration practices of citizens from specific countries of origin.
Switzerland	Asylum Applications	Four possible scenarios are decided by experts, in terms of the number of applicants: Low, Medium, High, and Very High. Each scenario considers the previous year’s number of asylum applications, changes in the countries of origin of asylum seekers, migration routes to and within Europe, and significant transit and/or target countries. Additionally, changes in the asylum and return practices of Switzerland and other European countries (including their implementation of the Dublin Convention) and the United States are also taken into account.
Australia	Labour migration	Targets are defined through expert elicitation, with consultations (e.g. with federal administration, State and Territories administration, researchers, modellers from the Department of Treasury).
Japan	Labour migration	Macro‑econometric model (forecasting future GDP etc., reflecting innovation & demographic changes) to estimate Japan’s foreign labour demand and country of origin’s labour supply, applied general equilibrium model to estimate industry-specific foreign labour demand (see Box 8.2).
Korea	Labour migration	Occupational labour supply projection and general equilibrium model forecasting labour demand (based on productivity and technological change), under 4 scenarios with different substitutability assumptions. Annual visa quota by sector can be established based on anticipated differential impacts of these labour demand/supply forecasts (see Box 8.2).
Poland	Labour migration	Quantitative model based on historical data augmented by qualitative research.
Türkiye	Labour migration	Delphi survey method using voluntary representation from public institutions, international organisations, academia and social partners.
Canada	All categories of regulated flows	Canada conducts demand forecasting for Temporary Residents (TR), Biometrics Enrollments, Citizenship Grants and Proofs, Passport Demand, and Permanent Resident Cards (Phase II). The forecasting framework for TR and Biometrics includes key categories such as Temporary Resident Visas (TRV), Work Permits (WP), Study Permits (SP), and their associated extensions. Forecasts are generated monthly using ARIMAX, Exponential Smoothing, and Unobserved Components models, supported by operational insights, domain expertise, and business knowledge. Assumption-based forecasting is also applied during periods of uncertainty, such as pandemics, conflicts, or significant policy changes. Building on this foundation, IRCC plans to explore the use of machine learning and hybrid statistical – machine learning approaches for migration forecasting.
France	All categories of regulated flows	ARIMA and SARIMA models (depending of the migration category), using past data from 2011 (excluding 2020).
Norway	All categories of regulated flows	Regression models based on recent past data.
Finland	All categories of flows (but family migration)	Qualitative and quantitative models using expert opinion, operating environment data and factors driving change.
Sweden	All categories of flows	Mixed approach based on quantitative time series analysis of historical patterns of applications in Sweden, including predictive factors, complemented with expert-based qualitative analysis.
United Kingdom	All categories of flows	Stable migration scenario based on 2022 visas, stay rates and planned policy changes.
United States CIS	All categories of flows	Time‑series modelling selecting from a variety of time‑series forecasts (ETS, ARIMA, ARFIMA, STL, TBATS) to choose the best model fit, excluding outliers when external forces (e.g. fee changes) inorganically affect receipts. For family migration (Box 6.1), for instance they leverage structural immigration relationships to forecast the time of the application (e.g. filing the family reunification form two years after the initial form is approved).Results are then discussed with subject matter experts from the operational directorates and the office responsible for new regulations.

Checklist:

• Is the model considered fit for purpose from an operational perspective?

  • Model choice depends on the type of migration to be forecast (notably, its volatility) and on the time horizon concerned.

• Are the data adapted for this type of model?

  • Model choice depends on data frequency and quality, as well as whether past time series were stationary.

What resources do countries need to run forecasting?

Establishing a functional and sustainable migration forecasting system requires not only technical infrastructure and skilled personnel but also strong institutional co‑ordination, leadership, and long-term commitment across stakeholders.

Even the simplest forecasting approaches rely on solid IT infrastructure. Handling large and often complex datasets – such as migration flows, drivers, and contextual indicators – requires both sufficient storage and processing power. At a minimum, countries need a secure data server to store harmonised datasets, a computing environment capable of running statistical or machine learning models without performance delays, and a scraping server or API setup to ingest external data sources efficiently for big data or automated input systems. While such infrastructure is common in statistical agencies and research institutions, it is often less available in public administrations not typically focussed on data‑intensive work. The initial investment may seem high, but it lays the foundation for all subsequent forecasting work.

Box 4.4. Staffing Situation for Migration Forecasting in MAP Task Force countries

In the United States, as of April 2025, USCIS (U.S. Citizenship and Immigration Services) employed the equivalent of three full-time equivalents (FTEs) dedicated to forecasting, including two federal employees and one contractor. However, the standard staffing includes one branch chief and three peer contributors (all federal).

Furthermore, the Migration Analysis Center – the research hub located within the Office of Homeland Security Statistics (OHSS), which focusses on Monthly Encounter Projections relies on one FTE internally to manage projections. Additionally, the equivalent of 2 FTEs worth of expertise is contributed by subject matter experts who participate in surveys, monthly forecasting meetings, and other consultations.

In Switzerland, the migration analysis unit is made up of three analysts and a branch chief – although forecasting is only a small part of their activity. Considering the time spent by the staff for observation, analysis and explanation of trends in irregular migration and asylum, as well as the time spent on discussion between them and their country-of-origin expert counterparts, it is estimated that a total of 1.5 FTE is involved in asylum forecasting.

In the Netherlands, asylum forecasting is managed by a senior advisor, half-time employed by the Ministry of Justice. To produce the forecasts, the senior advisor uses migration related developments, which are analysed by a “Migration Radar” unit, comprising two FTEs. The senior advisor also uses expert intuitions gathered from colleagues in international meetings. The time of these colleagues cannot be quantified. About two to three staff in the department then calculate the required staffing and budgets for the asylum process, based on the forecasts. All in all, about 5 FTEs in the Netherlands have their activity linked with asylum forecasting, with roughly half on the frontend.

The Swedish Migration Agency (SMA) has a dedicated function within a unit that works on forecasting. The Unit for Migration Analysis has five full-time Migration Intelligence Analysts and one Senior Analyst. The Migration Intelligence Analysts analyse how domestic and foreign developments will impact the future number of first-time applications (e.g. asylum, work permits, study permits, family reunification) lodged at the SMA. The majority of their monitoring and analytical reports have an element of forecasting, which goes into an integrated model. The SMA produces a weekly situational report. The Unit for Migration Analysis also produces a quarterly prognosis paper – not directly connected to the situational reports – in co‑operation with the SMA’s Planning Department. Although the unit produces a number of other reports not directly connected to forecasting, the Planning Department co‑ordinates the SMA’s planning as well as financial and statistical reporting. Forecasts of expenditure and operational forecasts are made by the Planning Department in the quarterly prognosis paper. Three people in the department hold primary responsibility for co‑ordinating the prognosis report and developing and maintaining the forecast models: a Financial Controller, a Statistician, and an Operational Analyst. Most other experts and financial controllers who participate in this work do so as a minor part of their duties. In total, 9 staff are involved in migration forecast reporting in Sweden, but it is difficult to quantify exactly their full-time equivalent.

A dedicated core team is also essential, even for basic forecasting models (Box 4.4). This typically includes data analysts, statisticians, economists or coders who can manage data preparation, model development, and interpret results. As forecasting methods evolve, expertise in machine learning, web scraping, and big data analytics is increasingly valuable. However, these profiles are highly sought after in the private sector and often difficult to attract to public administrations due to wage constraints or hiring restrictions – especially in contexts where civil service positions are limited to nationals. In addition to technical staff, effective forecasting systems require the involvement of wider working groups across relevant ministries, statistical offices and policy units. These groups contribute to framing scenarios, validating assumptions, and discussing timeframes and outputs. Regular co‑ordination meetings are crucial for aligning expectations and ensuring the forecasts are actionable and policy-relevant.

Beyond staff and IT tools, several enabling conditions determine whether forecasting efforts will be effective and sustainable:

• Leadership and governance: A clear mandate and leadership structure are needed to champion the project and secure its integration into policymaking.

• Data access and sharing: Forecasting depends on the willingness of different actors – national statistical offices, ministries, border agencies, and international partners – to share timely and reliable data. Long-term data-sharing agreements, based on trust and mutual benefit, are critical.

• Capacity to identify and harness relevant data: Technical expertise must be complemented by the ability to map available data sources, assess quality and integrate them into forecasting models.

• Creating the right conditions: Adequate time, cross-institutional engagement, and political support must be in place to move from pilot projects to regularised forecasting systems.

Countries may also consider outsourcing parts of the forecasting work to national statistical offices, academic institutions, think tanks, or private sector firms – particularly when internal resources are limited. Outsourcing can offer access to cutting-edge technical expertise, but requires clear governance, quality control, and sustainability plans. Strategic partnerships can also foster innovation and enable joint access to restricted datasets or proprietary tools.

While most forecasting software tools (such as R or Python) are free and open-source, other costs may arise. Accessing data from commercial providers – such as airlines, social media platforms, or polling institutes – can be expensive. Some datasets produced by research projects or international organisations may only be free for academic use, excluding government users. In such cases, data access agreements, participation in expert networks, or financial contributions may be required.

Checklist:

• Can I secure medium- to long-term financial and human resources needed to develop and maintain a forecasting model?

  • Investing in IT infrastructure and staff with the relevant skills (internally or outsourced) is the foundation of an efficient forecasting system, which can only be maintained with the right staff mandate and political support.

• Is access to public and private data ensured overtime?

  • Data access agreements, participation in expert networks and financial contributions may be necessary to use data in the long run.

How complex should the model be?

In deciding between simple and complex models, the question is how close should it be to reality? Existing literature on simple versus complex demographic forecasts, dating back to the 1990s, has been largely inconclusive about which is the better choice (e.g. Ahlburg (1995[39])). Similarly, in the area of general forecasting, a series of “Modelling Competitions” organised by Makridakis and co‑authors since the 1980s (e.g. Makridakis, Spiliotis and Assimakopoulos (2022[40])), yielded conclusions that varied and changed over time. Earlier competitions favoured simple models, such as exponential smoothing, while the most recent editions indicated increasing accuracy of complex, and often proprietary, machine learning algorithms, sometimes coupled with statistical methods and other approaches, such as forecast averaging (Makridakis, Spiliotis and Assimakopoulos (2022[40])). This clearly reflects the rapid evolution of the machine learning algorithms, availability of high-granularity data, and exponential increase in computational power in the last decades.

By design, the competitive advantage of machine learning and hybrid methods grows with the size of the available datasets and the frequency of data. For regular migration flows with quarterly or annual data, such as labour, study or family migration, we can expect simpler models to perform nearly as well as more complex ones, while being easier to implement, maintain and explain to stakeholders. In such situations, it makes sense to choose the lower complexity models: when performance is similar, a simpler model is likely to be a more cost-effective. At the same time, for more volatile migration processes, such as irregular border crossings or asylum, for which several large sources of high-frequency data exist, more sophisticated methods may be justified.

Another dimension of modelling complexity concerns the representation of the mechanisms driving migration. Ideally, models should include a causal description of underlying processes, as it can significantly enhance predictive capabilities (Willekens (2019[11]); see also Chapter 9). In recent years, there have been several attempts to model the causal dynamics of migration using agent-based models. In these models, simulated agents (such as migrants and decision makers) make choices based on their individual situations, knowledge, interactions with others and environment (e.g. Kniveton, Smith and Wood (2011[41]); Naivinit et al. (2010[42]); Klabunde and Willekens (2016[12]); Suleimenova and Groen (2020[43]); Bijak (2022[44]); Hinsch and Bijak (2023[45])). Whether the resulting model uncertainty is small enough to enable meaningful predictions, is another open question, addressed in Chapter 7. Still, for most practical applications, such complex models are too idiosyncratic to enable efficient implementation at a scale needed for most of the challenges of migration policy and practice.

Checklist:

• Is the model foreseen fit for purpose from a policy perspective?

  • Choosing the right model requires ensuring that the uncertainty is not too large to estimate the impact of relevant policy changes.

• What would be the pros and cons for using a simpler approach/model?

  • Model complexity depends on the migration category considered, as the level of uncertainty and the availability of high-frequency data varies. Forecasting volatile forced migration may justify use of sophisticated methods, while simpler methods usually work for regulated migration flows.

How often should model results be updated?

No predictive model of social processes, no matter how complex and how accurate, can indefinitely withstand the test of time. One of the ways in which the forecasting approaches can be “future‑proofed” – made more resilient and more reactive to changing migration reality – is to update them regularly. The update mechanism needs to be a part of the “business as usual” operations, built by design into processes aimed at ensuring anticipation and preparedness. Updating a model does not imply that the original model has failed; rather, updates reflect the availability of new data, improved methodologies, or changing contextual factors, which may affect the relevance or reliability of earlier results. Regular updates help ensure that models remain aligned with current realities and continue to support sound decision making.

The frequency of update clearly depends in part on the availability of resources. From the methodological viewpoint, the key variables are the horizon of the predictions and the frequency of data – whether they are daily, weekly, monthly, quarterly, or annual. For predictions with a horizon of 12‑18 months, a standard routine of annual updates seems appropriate: a full new year of data usually allows for recalibrating not only the trends, but also any patterns that may exhibit seasonality. At the same time, early warning models (for asylum or irregular migration forecasts) may need to be updated at least monthly, so that they can serve their intended purposes. Similarly, for long-range migration scenarios, such as those required for population projections, updates every couple of years would suffice, in line with the current practice of many national and international statistical institutions (UN, 2024[46]).

As for data frequency, methods for integrating data of different frequencies into migration forecasts exist, for example relying on multivariate models (such as vector autoregressions – VAR, e.g. Barker and Bijak (2025[47]). Using such models comes at an additional expense of increased predictive uncertainty, and additional resources needed to carry out the updates. Still, the extra effort may be justified if some covariates of the forecasting models are available at higher frequency than the migration variables (for instance, media reporting of events of political violence may be available daily, while asylum applications numbers may only be collated at a lower frequency). There may be some additional signal in such data. Automating data acquisition (for example via dedicated API – Application Programming Interface, or by web scraping) and re‑estimation of forecasting models can help implement the update routines in practice.

In addition to regular updates, the forecasting models will also need amending and resetting outside of the standard cycle, whenever there is a major migration event, change in political circumstances, or the policy environment. This would be a simple reflection of the dynamic nature of migration, moving from one social, political and economic equilibrium to the next often in a matter of months, if not days. Examples of such events that would trigger updates can include an eruption of major conflict leading to a large number of asylum seekers moving to other countries, a major shift in a visa regime, either relaxing or tightening the rules of admission and stay, or an economic downturn, likely to drive either more or less labour migration, depending on the nature of a particular event.

Checklist:

• Can routine updates of the data in the model be matched with the frequency of the prediction?

  • Updates are necessary but only possible as new data become available or when a major event occurs.

References

[39] Ahlburg, D. (1995), “Simple versus complex models: Evaluation, accuracy, and combining”, Mathematical Population Studies, Vol. 5/3, pp. 281-290, https://doi.org/10.1080/08898489509525406.

[26] Alexander, M., K. Polimis and E. Zagheni (2019), “The Impact of Hurricane Maria on Out‐migration from Puerto Rico: Evidence from Facebook Data”, Population and Development Review, Vol. 45/3, pp. 617-630, https://doi.org/10.1111/padr.12289.

[31] Avramescu, A. and A. Wisniowski (2021), “Now-casting Romanian migration into the United Kingdom by using Google Search engine data”, Demographic Research, Vol. 45, pp. 1219-1254, https://doi.org/10.4054/demres.2021.45.40.

[23] Azose, J. and A. Raftery (2015), “Bayesian Probabilistic Projection of International Migration”, Demography, Vol. 52/5, pp. 1627-1650, https://doi.org/10.1007/s13524-015-0415-0.

[24] Azose, J., H. Sevcikova and A. Raftery (2023), “bayesMig: Bayesian Projection of Migration”, CRAN: Contributed Packages (database), https://doi.org/10.32614/cran.package.bayesmig (accessed on 4 September 2025).

[47] Barker, E. and J. Bijak (2025), “Mixed-frequency VAR: a new approach to forecasting migration in Europe using macroeconomic data”, Data & Policy, Vol. 7, https://doi.org/10.1017/dap.2024.82.

[37] Barker, E. and J. Bijak (2022), Could we have seen it coming? Towards an early warning system for asylum applications in the EU..

[14] Beine, M., S. Bertoli and J. Fernández‐Huertas Moraga (2015), “A Practitioners’ Guide to Gravity Models of International Migration”, The World Economy, Vol. 39/4, pp. 496-512, https://doi.org/10.1111/twec.12265.

[15] Beyer, R., J. Schewe and H. Lotze-Campen (2022), “Gravity models do not explain, and cannot predict, international migration dynamics”, Humanities and Social Sciences Communications, Vol. 9/1, https://doi.org/10.1057/s41599-022-01067-x.

[35] Bijak, J. (ed.) (2024), From Uncertainty to Policy: A Guide to Migration Scenarios, Edward Elgar Publishing, https://doi.org/10.4337/9781035319800.

[44] Bijak, J. (2022), Towards Bayesian Model-Based Demography, Springer International Publishing, Cham, https://doi.org/10.1007/978-3-030-83039-7.

[1] Bijak, J. (2011), Forecasting International Migration in Europe: A Bayesian View, Springer Netherlands, Dordrecht, https://doi.org/10.1007/978-90-481-8897-0.

[20] Bijak, J. (2005), “Bayesian Methods in International Migration Forecasting”, CEFMR Working Paper.

[5] Bijak, J. et al. (2019), “Assessing time series models for forecasting international migration: Lessons from the United Kingdom”, Journal of Forecasting, Vol. 38/5, pp. 470-487, https://doi.org/10.1002/for.2576.

[19] Bijak, J. and A. Wiśniowski (2010), “Bayesian Forecasting of Immigration to Selected European Countries by using Expert Knowledge”, Journal of the Royal Statistical Society Series A: Statistics in Society, Vol. 173/4, pp. 775-796, https://doi.org/10.1111/j.1467-985x.2009.00635.x.

[30] Böhme, M., A. Gröger and T. Stöhr (2020), “Searching for a better life: Predicting international migration with online search keywords”, Journal of Development Economics, Vol. 142, p. 102347, https://doi.org/10.1016/j.jdeveco.2019.04.002.

[17] Burzynski, M., C. Deuster and F. Docquier (2020), “Geography of skills and global inequality”, Journal of Development Economics, Vol. 142, p. 102333, https://doi.org/10.1016/j.jdeveco.2019.02.003.

[21] Carammia, M., S. Iacus and T. Wilkin (2022), “Forecasting asylum-related migration flows with machine learning and data at scale”, Scientific Reports, Vol. 12/1, https://doi.org/10.1038/s41598-022-05241-8.

[27] Culora, A. et al. (2021), Using social media data to ’nowcast’ international migration around the globe..

[18] Dao, T. et al. (2021), “Global migration in the twentieth and twenty-first centuries: the unstoppable force of demography”, Review of World Economics, Vol. 157/2, pp. 417-449, https://doi.org/10.1007/s10290-020-00402-1.

[8] de Beer, J. (1997), “The effect of uncertainty of migration on national population forecasts: The case of the netherlands”, Journal of Official Statistics, Vol. 13/3.

[4] Disney, G. et al. (2015), Evaluation of existing migration forecasting methods and models.

[45] Hinsch, M. and J. Bijak (2023), “The Effects of Information on the Formation of Migration Routes and the Dynamics of Migration”, Artificial Life, Vol. 29/1, pp. 3-20, https://doi.org/10.1162/artl_a_00388.

[33] IDMC (2022), Nowcasting internal migration movements by using an AI-enabled news analyses tool. Internal Displacement Monitoring Centre.

[9] Keilman, N., D. Pham and A. Hetland (2001), Norway’s Uncertain Demographic Future, http://hdl.handle.net/11250/181350.

[12] Klabunde, A. and F. Willekens (2016), “Decision-Making in Agent-Based Models of Migration: State of the Art and Challenges”, European Journal of Population, Vol. 32/1, pp. 73-97, https://doi.org/10.1007/s10680-015-9362-0.

[41] Kniveton, D., C. Smith and S. Wood (2011), “Agent-based model simulations of future changes in migration flows for Burkina Faso”, Global Environmental Change, Vol. 21, pp. S34-S40, https://doi.org/10.1016/j.gloenvcha.2011.09.006.

[25] Lai, S. et al. (2019), “Exploring the use of mobile phone data for national migration statistics”, Palgrave Communications, Vol. 5/1, https://doi.org/10.1057/s41599-019-0242-9.

[28] Leasure, D. et al. (2023), “Nowcasting Daily Population Displacement in Ukraine through Social Media Advertising Data”, Population and Development Review, Vol. 49/2, pp. 231-254, https://doi.org/10.1111/padr.12558.

[10] Lutz, W. et al. (2000), New Developments in the Methodology of Expert- and Argument-Based Probabilistic Population Forecasting.

[6] Lutz, W. and S. Scherbov (1998), , European Journal of Population, Vol. 14/1, pp. 1-17, https://doi.org/10.1023/a:1006040321755.

[40] Makridakis, S., E. Spiliotis and V. Assimakopoulos (2022), “M5 accuracy competition: Results, findings, and conclusions”, International Journal of Forecasting, Vol. 38/4, pp. 1346-1364, https://doi.org/10.1016/j.ijforecast.2021.11.013.

[13] McAlpine, A. et al. (2020), “Agent-based modeling for migration and modern slavery research: a systematic review”, Journal of Computational Social Science, Vol. 4/1, pp. 243-332, https://doi.org/10.1007/s42001-020-00076-7.

[32] Morgenstern, S. and O. Strijbis (2024), “Forecasting migration movements using prediction markets”, Comparative Migration Studies, Vol. 12/1, https://doi.org/10.1186/s40878-024-00404-0.

[42] Naivinit, W. et al. (2010), “Participatory agent-based modeling and simulation of rice production and labor migrations in Northeast Thailand”, Environmental Modelling & Software, Vol. 25/11, pp. 1345-1358, https://doi.org/10.1016/j.envsoft.2010.01.012.

[36] Napierała, J. et al. (2021), “Toward an Early Warning System for Monitoring Asylum-Related Migration Flows in Europe”, International Migration Review, Vol. 56/1, pp. 33-62, https://doi.org/10.1177/01979183211035736.

[29] Rampazzo, F. et al. (2024), “Assessing Timely Migration Trends Through Digital Traces: A Case Study of the UK Before Brexit”, International Migration Review, Vol. 59/1, pp. 119-140, https://doi.org/10.1177/01979183241247009.

[38] Raymer, J. and F. Willekens (eds.) (2008), Forecasting international migration: time series projections vs argument-based, John Wiley.

[22] Robinson, C. and B. Dilkina (2018), “A Machine Learning Approach to Modeling Human Migration”, Proceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable Societies, pp. 1-8, https://doi.org/10.1145/3209811.3209868.

[2] Sardoschau, S. (2020), The Future of Migration to Germany.“ Assessing Methods in Migration Forecasting.

[7] Sohst, R., E. Acostamadiedo and J. Tjaden (2023), “Reducing uncertainty in Delphi surveys: A case study on immigration to the EU”, Demographic Research, Vol. 49, pp. 983-1020, https://doi.org/10.4054/demres.2023.49.36.

[43] Suleimenova, D. and D. Groen (2020), “How Policy Decisions Affect Refugee Journeys in South Sudan: A Study Using Automated Ensemble Simulations”, Journal of Artificial Societies and Social Simulation, Vol. 23/1, https://doi.org/10.18564/jasss.4193.

[46] UN (2024), World Population Prospects – 2024 Revision, https://population.un.org/wpp (accessed on 15 November 2024).

[34] UNHCR (2025), Nowcasting of refugee and asylum-seeker statistics.

[16] Welch, N. and A. Raftery (2022), “Probabilistic forecasts of international bilateral migration flows”, Proceedings of the National Academy of Sciences, Vol. 119/35, https://doi.org/10.1073/pnas.2203822119.

[11] Willekens, F. (2019), “Towards causal forecasting of international migration”, Vienna Yearbook of Population Research, Vol. 1, pp. 199-218, https://doi.org/10.1553/populationyearbook2018s199.

[3] Wiśniowski, A. (2021), “Migration forecasting using new technology and methods”, in Research Handbook on International Migration and Digital Technology, Edward Elgar Publishing, https://doi.org/10.4337/9781839100611.00039.