Long-Memory Models Beat Neural Networks in Forecasting Somalia’s Pastoral Red Meat Decline

Somalia’s pastoral livestock economy, which supplies a livelihood for more than 60 percent of the population and roughly 80 percent of agricultural GDP, is heading toward a slow but measurable plateau, according to a new forecasting study published in the journal Pastoralism. Researchers from Beder International University in Hargeisa and Amoud University in Borama assembled 64 years of production data, tested seven competing forecasting frameworks against each other, and produced the first head-to-head comparison of classical statistics, state-space models, long-memory econometrics, and neural networks ever applied to Somalia’s red meat sector. Their verdict: a decades-old statistical technique built for long-range persistence outperformed everything else, including modern machine learning.

The stakes extend far beyond academic bragging rights. Livestock herding in Somalia contributes about 45 percent of national GDP and more than 80 percent of foreign currency earnings, largely through exports to the Gulf States. Yet the sector is battered by recurrent droughts, transboundary animal disease outbreaks, and sudden import bans from destination markets. When shocks strike, their effects do not fade quickly; herd rebuilding takes years, and the losses ripple through national meat output for a decade or more. Accurate forecasting is therefore not a luxury but a foundation for food security planning in one of the world’s most climate-vulnerable states.

The research team, led by Abdirisak Osman Bade, drew on the Food and Agriculture Organization’s FAOSTAT database, filtering the raw records to include only the four pillars of the pastoral economy: camels, cattle, sheep, and goats. Commercial poultry was deliberately excluded because it plays almost no role in Somalia’s rangeland-based production system. After cleaning, the dataset yielded a continuous series of 64 annual observations from 1961 to 2024, with mean production of 144,214 metric tons per year and a standard deviation of roughly 40,269 tons. The distribution was nearly symmetric, with a skewness of just −0.15, but its flat kurtosis of −1.52 hinted at the long stretches of stable output punctuated by crisis-driven collapses that define the country’s pastoral history.

Before any model could be trusted, the team had to establish the statistical character of the series. Augmented Dickey-Fuller and Phillips-Perron unit-root tests both failed to reject nonstationarity in the raw data, with p-values of 0.544 and 0.424 respectively. After first differencing, however, both tests rejected the null decisively at p < 0.001, confirming that production is integrated of order one, or I(1). A Zivot-Andrews test allowing for structural breaks in both intercept and trend returned a statistic of −4.3992 and pinpointed a breakpoint around the year 2000, reinforcing the picture of a series shaped by regime shifts rather than smooth evolution.

Seven forecasting frameworks then competed on a blind out-of-sample test covering 2015 to 2024, having been trained on data from 1961 to 2014. The contenders included the classical ARIMA and Theta models, the state-space approaches ETS, BATS, and TBATS, the fractionally integrated ARFIMA model, and an autoregressive neural network. The ARFIMA model, which permits fractional differencing and thereby captures hyperbolically decaying autocorrelations, dominated every accuracy metric. It achieved a symmetric mean absolute percentage error of just 3.11 percent, a root mean squared error of 10,836.89 tons, a mean absolute scaled error of 0.85, and a Theil’s U statistic of 0.69, well below the 1.0 threshold that marks superiority over a naive random walk.

The neural network came second, with an sMAPE of 4.49 percent and an RMSE of 12,825.75, confirming that machine learning can capture nonlinear patterns but cannot match long-memory econometrics when the historical record is short and volatile. The classical and state-space models fared far worse. ARIMA posted an sMAPE of 11.41 percent, while ETS reached 12.20 percent, and BATS and TBATS produced identical errors because Somalia’s annual data contains no seasonality for the trigonometric components of TBATS to exploit. The authors attribute these failures to systematic bias: short-memory models produce overly smoothed trajectories that completely miss real-world shocks, assuming that disturbances either decay rapidly or persist indefinitely, when in fact drought impacts linger for years at intermediate strength.

Rolling-origin cross-validation reinforced the ranking. Using an expanding-window procedure that repeatedly re-estimated each model across sequential training partitions, the ARFIMA model posted an aggregate RMSE of 10,595.91 against 11,900.06 for the neural network, demonstrating that its dominance was not an artifact of one particular test decade. The biological explanation lies in the slow recovery of herd demographics. The 2016/17 drought alone caused an estimated deficit of 14.8 million animal births, including 10.5 million among goats and 4 million among sheep, alongside 4 million excess deaths. Such reproductive gaps translate into multi-year production lags that only fractional integration can represent mathematically.

The species-level breakdown reveals a striking ecological asymmetry. Camel meat production rose steadily across the full 64-year horizon, barely dipping during droughts, a testament to the species’ ability to maintain milk and meat offtake under extreme water scarcity. Cattle production, by contrast, collapsed dramatically during the 1991 civil conflict and the multi-season drought of 2022, while goats and sheep remained comparatively stable, with goats showing the stronger post-drought recovery. Local surveys of drought-stricken districts in southwest Somalia report herd losses of up to 62.1 percent for goats, 68.6 percent for cattle, 74.7 percent for sheep, and 63.5 percent for camels, figures that align closely with the contraction the models now project.

Fitted to the full historical record, the ARFIMA model projects a slow, steady decline in national red meat output through 2034, falling from 175,332.80 tons in 2025 to 168,884.30 tons by 2034. Confidence bands widen with the forecast horizon, stretching from roughly 113,000 to nearly 225,000 tons at the 95 percent level by the end of the decade. The authors interpret this projected plateau as evidence that Somalia’s traditional pastoral system may be approaching its dynamic ecological and structural limits, as chronic pasture depletion, reduced biomass regeneration, and compounding droughts constrain the demographic expansion of herds toward a new equilibrium.

The findings carry a pointed message for policymakers and donors. The study argues that transitioning from reactive, crisis-driven aid to proactive climate resilience requires targeted investment in digital index-based livestock insurance, commercial fodder value chains, and mobile veterinary infrastructure, all anchored to mathematically validated baselines of the kind the ARFIMA model now provides. The authors also acknowledge limitations: the univariate design cannot incorporate climate covariates such as vegetation indices or rainfall anomalies, the small sample restricted the neural network’s architectural depth, and FAOSTAT’s national aggregations inevitably blend informal cross-border trade and unrecorded subsistence slaughter into the official figures. Future work, they suggest, should integrate high-resolution spatial and climatic data within hybrid architectures. For now, the lesson is clear: in fragile dryland economies, the memory of past shocks is long, and only models that remember can see what is coming.

Subject of Research: Comparative time-series forecasting of pastoralist red meat production in Somalia using classical, state-space, long-memory, and neural network models

Article Title: Forecasting pastoralist red meat production trajectories in a fragile and climate-vulnerable state: a comparative evaluation of classical, state-space, long-memory, and neural network models in Somalia

Article References: Bade, A. O., Shire, D. M., Hassan, S. M., & Muse, A. H. (2026). Forecasting pastoralist red meat production trajectories in a fragile and climate-vulnerable state: a comparative evaluation of classical, state-space, long-memory, and neural network models in Somalia. Pastoralism: Research, Policy and Practice, 16, Article 17161. https://doi.org/10.3389/past.2026.17161

Image Credits: AI Generated

DOI: 10.3389/past.2026.17161

Keywords: Somalia, pastoralism, livestock forecasting, ARFIMA, long memory, neural networks, ARIMA, climate vulnerability, food security, drought, time series analysis, FAOSTAT