Forecasting sisal export value in Tanzania: a comparison of SARIMA, machine learning, and hybrid models

Tanzania’s sisal exports are an important source of foreign exchange and rural income, but their value is volatile, reflecting global demand cycles, competition from synthetic fibres, climatic variability and domestic supply constraints. Preliminary diagnostics shaped the modelling strategy: the series is seasonal, with a significant seasonal moving-average term at lag 12 ( $$\:\theta\:\text{\:=\:-0.1428}$$ , $$\:p\text{\:=\:0.003}$$ ) consistent with the annual harvesting and shipment cycle, and non-linear, exhibiting heteroskedasticity (H = 10.56, p < 0.01), excess kurtosis (11.85) and non-normal residuals (Jarque–Bera = 1172.98, p < 0.01) under a linear specification. Five models were estimated on 335 monthly observations from the Bank of Tanzania (May 1996–March 2024): SARIMA, Support Vector Regression (SVR), Long Short-Term Memory (LSTM), Hybrid SARIMA–SVR and Hybrid SARIMA–LSTM, evaluated on a 20% hold-out set by RMSE, MAE and MAPE. The LSTM was most accurate (MAE = 0.7238; RMSE = 0.8168; MAPE = 38.79%), ahead of SVR (41.44%), Hybrid SARIMA–LSTM (40.92%), SARIMA (40.95%) and Hybrid SARIMA–SVR (41.56%). Diebold–Mariano tests qualify this ranking: the LSTM significantly outperforms SVR (p = 0.0008) and SARIMA outperforms the Hybrid SARIMA–LSTM (p = 0.0014), but the LSTM advantage over SARIMA is not significant at the 5% level (p = 0.0522). Recursive 24-month forecasts trace a near-flat path of USD 2.28–2.42 million to March 2027 with wide bootstrapped intervals, the flatness reflecting mean reversion in recursive multi-step prediction rather than an expectation of stable exports. Machine learning is therefore competitive with, rather than superior to, a well-specified Box–Jenkins model on volatile commodity export data, and hybrid decomposition offers no gain where the linear stage leaves heteroskedastic, non-normal residuals. NA.

Authors

Institutions

Publication Details

Journal
Discover Analytics
Published
2026-11-19
DOI
https://doi.org/10.1007/s44257-026-00104-2
Primary Topic
Forecasting Techniques and Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Forecasting sisal export value in Tanzania: a comparison of SARIMA, machine learning, and hybrid models

Twahil Hemed Shakiru, Godfrey Joseph Saqware, Modest Stephen Dindili
Discover Analytics
Forecasting Techniques and Applications
article

Forecasting sisal export value in Tanzania: a comparison of SARIMA, machine learning, and hybrid models

Twahil Hemed Shakiru, Godfrey Joseph Saqware, Modest Stephen Dindili
article en

Abstract

Tanzania’s sisal exports are an important source of foreign exchange and rural income, but their value is volatile, reflecting global demand cycles, competition from synthetic fibres, climatic variability and domestic supply constraints. Preliminary diagnostics shaped the modelling strategy: the series is seasonal, with a significant seasonal moving-average term at lag 12 ( $$\:\theta\:\text{\:=\:-0.1428}$$ , $$\:p\text{\:=\:0.003}$$ ) consistent with the annual harvesting and shipment cycle, and non-linear, exhibiting heteroskedasticity (H = 10.56, p < 0.01), excess kurtosis (11.85) and non-normal residuals (Jarque–Bera = 1172.98, p < 0.01) under a linear specification. Five models were estimated on 335 monthly observations from the Bank of Tanzania (May 1996–March 2024): SARIMA, Support Vector Regression (SVR), Long Short-Term Memory (LSTM), Hybrid SARIMA–SVR and Hybrid SARIMA–LSTM, evaluated on a 20% hold-out set by RMSE, MAE and MAPE. The LSTM was most accurate (MAE = 0.7238; RMSE = 0.8168; MAPE = 38.79%), ahead of SVR (41.44%), Hybrid SARIMA–LSTM (40.92%), SARIMA (40.95%) and Hybrid SARIMA–SVR (41.56%). Diebold–Mariano tests qualify this ranking: the LSTM significantly outperforms SVR (p = 0.0008) and SARIMA outperforms the Hybrid SARIMA–LSTM (p = 0.0014), but the LSTM advantage over SARIMA is not significant at the 5% level (p = 0.0522). Recursive 24-month forecasts trace a near-flat path of USD 2.28–2.42 million to March 2027 with wide bootstrapped intervals, the flatness reflecting mean reversion in recursive multi-step prediction rather than an expectation of stable exports. Machine learning is therefore competitive with, rather than superior to, a well-specified Box–Jenkins model on volatile commodity export data, and hybrid decomposition offers no gain where the linear stage leaves heteroskedastic, non-normal residuals. NA.

Discover AnalyticsVol. 4(1)
University of Dar es Salaam (TZ)
Forecasting Techniques and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon