The nowcasting paradigm

The lag in ODA data publication is the kind of problem nowcasting was developed to address. The term, adopted by macroeconomists, describes the estimation of the present or recent past value of an official statistic [3], [4]. A nowcasting system estimates a lagged statistic by drawing on timely but imperfect data sources, known as signals, that are available before the official statistic is published. These signals carry partial information about the current value of the statistic. Central banks and statistical agencies routinely nowcast economic growth, inflation, trade, and employment statistics.

While ODA is not a macroeconomic indicator and macroeconomic nowcasting methods do not transfer directly, the underlying logic of nowcasting is relevant: estimating a lagged statistic from timely imperfect signals within a statistical framework. Three design principles should guide the construction of an ODA nowcasting system:[1]

1. It targets a specific future statistical release. In this case, the targeted release is the OECD’s annual Creditor Reporting System (CRS) publication.
2. Its estimates update mechanically as new information arrives, each revision reflecting all available information at the time of estimation.
3. It produces estimates with formal uncertainty bounds that narrow as signals accumulate, rather than unquantified point estimates or scenario-based projections.

The development community already produces ad hoc estimates, scenario based projections, and forecasts. A formal nowcasting system complements these efforts by providing an estimate that gets updated as new information arrives, with quantified uncertainty at each point in time.

ODA data

The CRS, maintained by the DAC, is the authoritative source on detailed ODA flows, providing activity-level information on donors, recipients, sectors, channels, and financial amounts, including commitments and disbursements. The CRS data is published annually. Preliminary donor-level data for the previous year is published in April and the full activity-level microdata is published in late December. This produces a substantial lag. An individual disbursement made in January of a given year will not appear in official statistics until the full CRS release in December of the following year, 23 months later.

ODA data has important characteristics that differentiate it from the traditional macroeconomic variables for which nowcasting methods were developed. Three characteristics are important to consider.

Aid inertia. Historically, ODA volumes have been highly persistent. For many donor-sector combinations, last year's ODA volumes are a strong predictor of this year's volumes. This persistence reflects an institutional reality: multi-year programme commitments, legislative earmarks, multilateral replenishment cycles, and bureaucratic processes create a baseline level of spending that tends to change gradually. Past ODA volumes are consistently found to be among the strongest determinants of current volumes [5]. This persistence, however, is not uniform. Humanitarian ODA is more volatile and crisis-responsive, and the strength of inertia may vary across ODA types. More broadly, the predictive power of aid inertia fails when there is a shock or crisis. That is precisely when an estimate is needed most, as the 2025 donor cuts illustrate. A system that only relies on past ODA levels will perform well during stable periods but fail during a crisis.

Shallow time series. Comparable and complete data at the donor, sector, and recipient level is only available from around 2002, providing roughly 22 years of comparable historical data.( 2 ) The usable window for some signals is even shorter. This is a short timeframe by macroeconomic standards—nowcasting models are typically trained on decades of monthly or quarterly data. That makes nowcasting approaches that require long time series for reliable estimation not viable for ODA.

Donor heterogeneity. ODA profiles differ in composition, volatility, and predictability. Some donors maintain relatively stable sectoral allocations year to year; others shift substantially in response to humanitarian emergencies or political priorities. In-donor refugee costs, in particular, have become a significant and volatile component of ODA for some donors, adding a further source of cross-donor variation. This variation matters because the top DAC donors account for a large share of total bilateral ODA. Correctly estimating totals for the largest donors is essential for aggregate accuracy. Granular estimates where the information gap can be significant, however, such as sector-specific flows, are valuable. A system must accommodate this variation rather than assuming a common specification fits all donors.

Given the CRS publication calendar and the characteristics of ODA data, we propose targeting bilateral gross ODA disbursements. Disbursements capture actual resource transfers rather than pledges, are smoother and more persistent than commitments, and are highly policy relevant. Gross disbursements are also less sensitive to changes in how ODA is defined and accounted for, providing a more stable variable to target.( 3 )

Annual estimates at the donor level are the natural level of detail from which to start. Select DAC donors account for the majority of ODA volumes, and an annual timeframe follows the publication cycle of the CRS. More granular estimates—by sector, recipient, and at sub-annual frequency—are where the informational value is greatest, but where feasibility depends on signal coverage and time series depth.( 4 )

ODA signals

The ODA data landscape contains a broader set of potential signals than is commonly employed in estimation work. Signals vary in what they measure, how frequently they arrive, and how directly they relate to the target variable. Some signals provide direct partial measurement of the variable, some bound the range of plausible estimates, and others reflect the broader environment in which ODA policy operates. We organise these signals into three categories—direct signals, constraint signals, and ambient signals—and identify potentially informative signals that should be evaluated.

Direct signals: Direct signals, or cell-specific signals, provide direct partial information about the target variable itself.( 5 ) These may include disbursements reported externally to the CRS. International Aid Transparency Initiative (IATI) transaction data is the primary example, providing partial-year visibility into activity that will eventually appear in the CRS. Other direct signals may include disbursement data published by multilateral institutions on their own timelines. This includes the World Bank, UN agencies, and regional development banks, though this data pertains to multilateral rather than bilateral flows.

Constraint signals: Constraint signals do not directly measure a flow, but bound the space of plausible outcomes. The most important are donor budget documents and parliamentary appropriations, which set the envelope within which disbursements likely fall. Budget information is an important signal for aggregate donor volumes, and is already used in estimation efforts.( 6 ) Replenishment schedules for multilateral funds operate similarly: a confirmed IDA replenishment or a pledged contribution to the Global Fund constrains the eventual bilateral CRS entries for those channels. Prior-year CRS commitments with defined durations and drawdown schedules bound the plausible range of disbursements in subsequent years, though commitments can be cancelled or restructured. The OECD's preliminary estimates and DAC survey responses provide authoritative but infrequent constraint signals, arriving once per year.

Ambient signals. Ambient signals affect the broad ODA environment but do not map directly to specific flows. Donor country macroeconomic conditions, including growth rates, fiscal balances, and exchange rates, shape the fiscal space within which ODA budgets are set. Political and event-based signals such as government changes, elections, and geopolitical shifts can create sudden discontinuities in ODA, as the 2025 USAID cuts illustrate. Humanitarian crises and refugee flows recorded through UNHCR data or asylum statistics may be informative about specific donors, sectors, and recipients, particularly in-donor refugee costs. Some recipient-side signals may also carry information; for example, changes in foreign reserves for recipients with heavy ODA reliance may reflect changing ODA flows.

Signal quality, timing, and relevance

Whether a signal is informative for nowcasting is a separate question that depends on its quality, timing, and empirical relationship to the target variable.

Signal quality varies considerably across donors and sources. IATI coverage and reporting quality vary considerably across publishers, and harmonising IATI and CRS data requires significant data engineering effort.( 7 ) In addition, significant reporting to IATI increased in the mid-2010s, shortening comparable IATI-CRS overlap on which to estimate IATI’s reliability as a CRS signal. Budget data presents similar challenges. Donor budget documents do not exist in standardised machine-readable formats across the DAC. Mapping budget lines to CRS sectors and recipients is therefore not straightforward and requires institutional knowledge.

Signals do not arrive on a regular schedule. IATI data accumulates continuously but unevenly across publishers, budget information arrives when parliaments act, OECD preliminary releases arrive once a year, and macroeconomic indicators follow their own calendars. At any given point, signals with varying publication dates exist, creating an uneven "jagged edge” to the information set.( 8 ) This irregularity is a structural feature of the ODA signal landscape and a constraint that any estimation approach must accommodate.

The signal taxonomy above describes the landscape of potentially informative signals, not a validated set of model inputs. Whether any given signal demonstrably improves on simple persistence as a predictor of current or prior-year CRS values is an empirical question. Some signals may add little beyond what past ODA already captures; others may prove valuable for some donors but not others.

An important distinction should also be drawn between determinants of ODA and signals useful for nowcasting. The aid literature identifies the importance of donor-side factors as determinants of ODA [5][6][7], but these findings largely concern longer term variation in forecasting exercises. A variable that explains long-run differences in donor generosity may tell us little about how ODA is changing in the present. Nowcasting requires signals that are informative about short-run departures from the recent baseline. Which signals earn their place in an estimation system requires empirical testing.

The nowcasting toolkit

The nowcasting literature offers a range of estimation frameworks. Not all are suited to ODA. The characteristics identified above—aid inertia, shallow time series, irregular signal arrival, and donor heterogeneity—impose constraints on viable approaches. We assess the most relevant frameworks against these constraints to identify which approaches warrant empirical investigation.

Baselines

Carry-forward and simple autoregressive models exploit aid inertia directly. They assume past ODA is the best predictor of current ODA. These models set a surprisingly high bar. For many donor-sector combinations, simply carrying forward last year's value is hard to beat. Any credible nowcasting system should run baselines alongside more complex approaches, ensuring that additional complexity earns its place only when it demonstrably outperforms simple persistence.

Baselines, however, do not incorporate additional signals and fail precisely when an estimate is most needed—when inertia breaks. The 2025 donor cuts illustrate this. The value of more sophisticated approaches lies in detecting when the baseline no longer holds.

Regression-based approaches

Regression models incorporate additional signals as predictors of ODA. Two established approaches are particularly relevant.

Bridge equations regress the target variable on timely indicators to estimate its unpublished value. The approach is intuitive and interpretable. Bridge equation logic is a core component of well-known nowcasting systems such as the Atlanta Fed’s GDPNow. Bridge equations, however, require that all inputs be aggregated to a common frequency before estimation. For ODA, this means annual frequency, which discards within-year information and provides no mechanism for updating the estimate as signals arrive throughout the year.

Mixed Data Sampling (MIDAS) regressions improve on this by allowing higher-frequency signals to enter the estimation at their native frequencies rather than aggregating to an annual frequency [8].( 9 ) Panel-pooling extensions are also relevant given the shallow time series [9]. MIDAS assumes signals arrive at regular intervals, such as on a monthly or quarterly basis. Many ODA signals do not: IATI data accumulates irregularly, budget information arrives when parliaments act, and preliminary releases are event-driven. The regularity MIDAS requires is not present in much of the ODA signal landscape.

Both approaches may be viable for donor-level estimation when the number of signals is manageable, and particularly for prior-year estimation where most signals have already accumulated.( 10 ) But neither can update estimates sequentially as new information arrives or handle the irregular timing of ODA signals without discarding information.

State-space models

State-space models approach the problem differently. Rather than regressing on a set of predictors, they treat the current value of ODA as an unobserved state that evolves over time and is imperfectly revealed through available signals. When new information arrives, the estimate is updated along with revised uncertainty bounds [3] [4].

The state-space formulation handles the jagged edge of ODA signals naturally. As IATI data accumulates, the estimate updates. If budget information has not yet been released, the estimate holds and the uncertainty bounds reflect the missing information. There is no need to aggregate signals to a common frequency or to wait for the slowest signal to arrive. Formal uncertainty bounds and the ability to trace revisions to specific signal arrivals are direct outputs of the framework.

In macroeconomic nowcasting, state-space models are most commonly implemented as dynamic factor models.( 11 ) This is the approach behind the nowcasting systems of the New York Federal Reserve and the European Central Bank. These models require large panels of high-frequency indicators and long time series for reliable estimation, neither of which ODA provides. The state-space formulation is not tied to this implementation, however. A simpler specification that links the available signals directly to the latent ODA state retains the framework’s main advantages while requiring less data than dynamic factor models.

The risk is that a state-space model requires the analyst to specify how each signal relates to the unobserved state. With shallow time series and signals of uncertain predictive value, the model structure may encode assumptions that the data cannot adequately test.

Hierarchical pooling and Bayesian estimation

Estimating a single donor or donor-sector combination with few annual observations is challenging to do accurately. Two techniques address this problem directly.

Hierarchical modelling pools information across donors and sectors. Rather than estimating each case independently, the model assumes that donors share common underlying patterns while still allowing individual differences, using the data to learn how much variation exists across the group. Donors with limited data are drawn towards the group average, borrowing strength from the broader panel. Donors with abundant data can deviate from the shared pattern. For example, the United States and Luxembourg are not the same estimation problem, but the model does not need to know that in advance; it learns how different each donor is.

Bayesian estimation addresses shallow time series differently. Rather than relying on the observed data alone, it starts with prior beliefs about plausible values, such as the typical persistence of ODA, and updates those beliefs as new data is observed. Some categories of prior information for ODA are strong—aid persistence is extensively documented, budget envelopes are legislated, replenishment schedules are contractual. Bayesian estimation incorporates this domain knowledge formally rather than leaving it implicit. Where prior information is limited, priors can be set as weakly informative, allowing the data to play a larger role.

These two techniques reinforce each other. The hierarchy borrows information across donors and sectors; Bayesian estimation combines that pooled information with domain knowledge in a coherent probabilistic framework. Together, they compensate for what the shallow ODA time series alone cannot support.

Machine learning approaches

Machine learning methods, such as gradient boosting, random forests, and penalised regressions, can capture nonlinear patterns that structural models miss. They have been applied to nowcasting problems in other domains. For ODA, the constraints are significant: the shallow time series limits the ability to reliably test model performance, and these methods do not naturally produce interpretable uncertainty bounds. The most promising role may be to identify which signals matter most for individual donors, rather than to serve as the primary estimation framework.

Towards a framework

The constraints of ODA considerably narrow the field of viable approaches and point to a coherent candidate framework. A state-space model would treat current ODA as an unobserved quantity and update its estimate each time a new signal arrives. A hierarchical structure would pool information across donors and sectors, learning from patterns across the panel to compensate for the challenges that short time series pose. Bayesian estimation would incorporate institutional knowledge, such as the persistence of aid or the constraints imposed by published budgets, rather than relying on data alone. Together, these components form a system that would produce estimates with formal uncertainty bounds while accommodating differences across donors.

Whether this framework can outperform simpler baseline approaches is an empirical question that depends on what the ODA signal landscape can support. The path forward is to start with simple baselines, empirically evaluate what the signal landscape can support, and test whether the additional structure of this framework demonstrably improves estimation. Timely ODA estimates already exist through budget analysis, scenario projections, and direct extraction from sources such as IATI. A formal nowcasting system would complement these efforts by providing a statistical framework that updates systematically and quantifies uncertainty as signals accumulate.

Open questions

Model feasibility and validation

The framework described above is a candidate for investigation, not a validated model. Whether it delivers useful estimates for ODA requires empirical testing. Bayesian estimation is computationally feasible with modern probabilistic programming tools, though the practical implementation challenges should be evaluated. Validation requires pseudo-real-time backtesting: holding out recent years of CRS data, estimating them using only the signals that would have been available at the time, and comparing the estimates against the eventually published values. With roughly 22 years of comparable CRS data and some signals available for a shorter window, the number of backtesting vintages is small. This may be sufficient to assess whether a given approach beats naive baselines but is unlikely to support precise comparisons between competing models.

Architecture and design choices

Donor-level estimation for major DAC donors is the appropriate first test, capturing the majority of ODA volumes with a manageable number of time series and a relatively rich signal landscape. Donor-level totals, however, are of limited direct use to many potential users who need sectoral and geographic breakdowns. The system's policy relevance depends on whether it can extend to finer disaggregation.

Whether donor-sector estimation is feasible depends on signal coverage at that level and on whether hierarchical pooling can sufficiently compensate for short per-cell series. Recipient-level estimation is the most ambitious extension and the hardest to support. Disaggregated estimates must also be consistent with known totals, such as the donor-level aggregates published in the April preliminary CRS release. Methods for reconciling hierarchical time series are well established [10] [11], but how reconciliation should be integrated into this framework is an open question. The temporal target also matters: estimating prior-year ODA before the December CRS release, when most signals have already arrived, is a substantially easier problem than estimating current-year ODA mid-year.

Signal infrastructure and governance

An operational nowcasting system requires substantial engineering effort beyond the signal challenges described above. Whether the signals identified in this paper can be reliably extracted, standardised, and updated at sufficient frequency in practice requires investigation. Tracking historical vintages of signal data is also essential for honest backtesting, but is difficult to reconstruct retrospectively. ONE Data's Development Finance Observatory is building the data infrastructure a nowcasting system would depend on, but the gap between curated data and model-ready signal inputs is substantial.( 12 )

A nowcasting system also embeds assumptions that require periodic review as conditions change. Those include assessing how persistent ODA is, how informative each signal is, and the potential size of shocks. Signal reliability needs to be reassessed as publishers change reporting behaviours. These are ongoing governance requirements, not one-off design choices.

Structural breaks and the limits of the system

A nowcasting system estimates what has already happened but has not yet been measured. It does not predict political shocks, budget freezes, or agency disruptions—nor should it attempt to. This is a fundamental boundary of the approach.

When a shock to ODA occurs, a nowcasting system would begin quantifying effects as they propagate through observable signals, providing estimates sooner than the CRS publication cycle. But it would not anticipate the shift before it occurs, and in a fast-moving crisis the estimates may lag until signal effects become observable.

Forecasting future ODA levels under various policy scenarios and producing projections based on expert judgement are valuable exercises. But they are distinct from estimating the current state. Clarity about which exercise is most needed should inform decisionmaking about whether to invest in building a nowcasting system.

Conclusion

This paper argues that a formal estimation system for ODA—informed by the nowcasting paradigm—could be useful and is within methodological reach. Such a system would not have predicted the 2025 cuts, but as signals emerged, it would have begun quantifying their shape, producing estimates with formal uncertainty that updated as new information became available.

The ODA data landscape presents genuine constraints: shallow time series, irregular signal arrival, donor heterogeneity, and signals whose predictive value is uncertain. Those constraints narrow the range of viable approaches, but do not rule them out. A Bayesian hierarchical state-space formulation addresses each of the constraints within an integrated framework. Whether it can outperform simple baselines that rely on aid persistence is an empirical question and depends on what the signal landscape can support.

Decisions about programme continuity, funding reallocations, and advocacy strategy cannot wait for official data to be published. A system that produces timely estimates with quantified uncertainty would give the development community a stronger basis for decisions it is already making.

The path from a candidate framework to a working system is empirical. Our next priorities are to inventory and assess the signals identified in this paper, and others suggested by the community, and to backtest baseline approaches at the donor level. Decisions about which methods to take further and how far to push disaggregation depend on what those steps reveal. ONE Data intends to develop a nowcasting system as part of the Development Finance Observatory, and we welcome feedback and contributions from the development finance community on useful signals to consider, on the analytical needs that would benefit most from timely estimates, and on the trade-offs across the technical options outlined in this paper.

Please contact us at [email protected]