Nowcasting outbreak case and death counts from open-source reporting: a validation on the 2026 Ebola (Bundibugyo) outbreak in the Democratic Republic of the Congo

Technical Summary: Nowcasting Outbreak Case and Death Counts from Open-Source Reporting

Problem Statement
Official outbreak reports, while authoritative, suffer from inherent publication delays. Between situation reports, response teams must rely on stale data (often 2–5 days old) or extrapolate the official trend forward. This extrapolation fails to capture recent developments, particularly during critical phase transitions where an outbreak accelerates or slows. In low-surveillance environments or during acute emerging outbreaks, official data may be intermittent or entirely absent for periods, leaving a "reporting gap" where decision-makers lack a current estimate of cumulative case and death counts. The study addresses whether open-source intelligence (OSINT) can bridge this gap to provide a more accurate, real-time estimate of the current cumulative burden than simple extrapolation of the official series.

Methodology
The study proposes a statistical nowcasting framework that reconstructs cumulative counts from open sources and projects them to the present, anchored to the latest available official figure. The method is validated on the 2026 Bundibugyo virus Ebola outbreak in the Democratic Republic of the Congo (DRC).

1. Data Sources: The reconstruction utilizes 2,205 open-source figures from 67 distinct sources (aggregators, international wires, local reporting) covering the DRC national series. Official situation reports (INSP/COUSP) serve as the benchmark but are excluded from the reconstruction input.
2. Source Reliability & Reconstruction:
   • Each source is assigned a reliability weight ($w_s$) by a source-assessment model that does not use the official series as input, ensuring no look-ahead bias.
   • For each date, the reconstructed value ($R_t$) is the reliability-weighted median of reported figures, robust to single erroneous reports.
   • The series is projected onto a monotone non-decreasing cone via isotonic regression to ensure cumulative counts do not decrease.
   • The reconstruction uses only reports published on or before the "origin" date, strictly adhering to a no-look-ahead vintage backtest protocol.
3. Nowcasting Algorithm:
   • The final estimate ($\hat{C}$) anchors to the most recent official figure ($O_{t^*}$) and scales it by the growth observed in the open-source reconstruction between the last official date ($t^*$) and the present ($t_0$):
     $$\hat{C} = O_{t^*} \times \frac{\tilde{R}_{t_0}}{\tilde{R}_{t^*}}$$
   • Where $\tilde{R}$ is the smoothed mean of the last three reporting-day values.
   • This approach isolates the growth from the open sources while retaining the level of the official anchor, canceling out constant multiplicative offsets between open and official reporting.
4. Comparative Baselines: The study compares the open-source nowcast against:
   • Official-trend extrapolation: Growing the last official figure forward using the compound daily growth rate of the official series over the preceding seven days.
   • Holt's exponential smoothing: A damped-trend variant applied to the official series.
   • Blend: A log-space combination of the official trend and open-source nowcast, weighted by the staleness of the official figure ($L$).

Key Results
The validation, conducted over a real-time (vintage) backtest spanning roughly ten weeks, demonstrates that the open-source nowcast significantly outperforms official-trend extrapolation when the official data is stale. The evaluation window covers 63 official days for cases (23 May to 3 August 2026) and 48 days for deaths (7 June to 3 August 2026).

• Performance by Lag:
  • Fresh Data (1–3 days): When the official figure is recent, simple extrapolation is comparable to or slightly better than the open-source nowcast.
  • Stale Data (4+ days): The open-source nowcast becomes materially more accurate.
    • Cases: At a 7-day lag, the median absolute percentage error (MAPE) drops from 12.6% (official trend) to 3.9% (open-source). The improvement is statistically significant ($p < 10^{-5}$) starting at a 4-day lag.
    • Deaths: At a 7-day lag, MAPE drops from 11.3% to 6.8%. The advantage becomes significant around a 6-day lag.
• Phase Transitions: The primary value of the method is concentrated at epidemic turning points. During periods of steady growth, both methods perform similarly. However, when the trend shifts (acceleration or deceleration), the official extrapolation fails to detect the change, resulting in high errors (e.g., 27.0% MAPE for cases at turning points). The open-source nowcast captures these shifts, reducing error to roughly half (13.6% MAPE).
• Uncertainty: The nowcast is near-unbiased (median error $\approx$ 0) but exhibits an asymmetric error distribution with a downside tail (under-counting) during fast-growth phases. A blended approach (combining official trend and open-source) narrows this tail, halving the 80% error interval in mid-horizons.

Significance and Claims
The paper claims that open-source reporting provides a viable mechanism to reduce the latency inherent in official outbreak monitoring, specifically when the official data is several days old.

• Data over Model: The study emphasizes that the improvement comes from data (access to recent open-source reports) rather than model sophistication. Adding complex mechanistic models (e.g., SIR/SEIR) or advanced time-series forecasting (e.g., Holt's method) to the stale official series does not recover the missing information contained in the reporting gap.
• Operational Utility: The method is designed for real-time deployment where the "operating gap" (the time between the last official report and the present) is the standard condition. It provides a calibrated, independent estimate of the current cumulative count, which is essential for resource allocation and risk assessment.
• Generalizability: The approach is pathogen-agnostic and applicable to any outbreak where case/death figures appear in open reporting, provided the reporting density is sufficient.
• Limitations: The authors explicitly note that the method validates a faster path to the official series, not necessarily a more accurate path to the true biological burden (as official counts themselves under-ascertain). The method is a nowcast of observed counts, not a mechanistic transmission model. The validation is limited to a single outbreak, and the statistical significance is treated with caution due to the autocorrelated nature of cumulative data.

In summary, the paper demonstrates that for acute outbreaks with reporting lags, a simple open-source reconstruction anchored to the latest official figure offers a robust, low-latency alternative to trend extrapolation, particularly during the critical moments when an outbreak's trajectory changes.