A Tunisian manufacturer reports 2.25 million TND in annual cost savings and lower emissions after replacing traditional forecasting with CNN‑LSTM models, which improved demand-forecast accuracy and operational efficiency; the finding comes from a single-firm before‑and‑after analysis and may not generalize where data fragmentation and skills shortages persist.

Main Finding

A hybrid CNN–LSTM forecasting system deployed in a Tunisian manufacturing firm materially improved demand forecast accuracy and produced measurable dual economic and environmental gains. Compared with the firm’s legacy methods, the CNN–LSTM raised demand-forecast accuracy (DFA) from 65% to 88% (a 23 percentage-point / ~35.4% relative increase), reduced annual operating costs by TND 2.25 million (15%), cut inventory holding costs by 30%, improved OTIF from 88% to 96%, and reduced carbon emissions by 15%. Organizational readiness (data infrastructure and workforce AI literacy) was a key mediator of these gains.

Key Points

• Forecasting performance
  • MAPE: ARIMA 35.0% → MLP 22.1% → CNN–LSTM 12.0%
  • RMSE: ARIMA 1.85 → MLP 1.15 → CNN–LSTM 0.55
  • DFA: ARIMA 65% → MLP 77.9% → CNN–LSTM 88%
  • CNN–LSTM outperformed classical statistical models especially during volatile periods (spikes, promotions).
• Operational impacts (post-implementation)
  • Annual operating costs: TND 15.00M → TND 12.75M (−15%; TND 2.25M saving).
  • Inventory holding costs: −30% (as % of total inventory).
  • Stockouts reduced (stockout rate fell by ~73.3%).
  • OTIF improved from 88% to 96%.
• Environmental impacts
  • Scope 2/3 carbon emissions down 15%.
  • Energy consumption per unit −12%.
• Organizational & adoption barriers
  • Fragmented data architecture, disconnected spreadsheets, limited historical digitization.
  • Low AI literacy among planners; initial distrust of “black-box” outputs.
  • High upfront investment costs and limited in-house analytics expertise.
• Theory/tests
  • H1a (data readiness → DFA) and H1b (AI literacy → DFA) supported qualitatively and as mediators.
  • H2 (higher DFA → better economic & environmental performance) supported by KPIs.

Data & Methods

• Research design
  • Mixed-methods quasi-experimental field study in one Tunisian manufacturing firm over 12 months with pre-, during-, and post-intervention phases.
  • Triangulation: operational KPI time series + 35 semi-structured interviews across roles.
• Data
  • Operational data from ERP modules and dashboards: 12 KPIs measured daily/weekly (sales, production, lead times, energy use, carbon emissions, material waste, inventory metrics).
  • Historical multiyear time series; data cleaning and reconciliation required (part-number harmonization, digitization of paper records).
• Model & training
  • Hybrid CNN–LSTM multivariate forecasting model: sliding time window inputs (sales, promotions, production indicators, external variables), 1D convolutional layer(s) for local pattern extraction, LSTM layer(s) for long-term temporal dependencies, dense output layer.
  • Preprocessing: Min–Max scaling.
  • Training: 80/20 train/test split, Adam optimizer, early stopping, hyperparameter tuning via Bayesian optimization.
  • Benchmarked against ARIMA, Holt–Winters, and MLP; metrics: MAPE, RMSE, DFA = (1 − MAPE)×100.
• Validity & limitations reported
  • Construct validity via triangulation; internal validity via pre/post comparisons; external validity cautioned—single-firm, context-specific to emerging-market constraints.
  • Limitations: single-case design, potential omitted unintended consequences (labor displacement, cybersecurity), scalability costs not fully modeled.

Implications for AI Economics

• Empirical link between forecast accuracy and firm-level economic and environmental outcomes
  • Quantifies how improved predictive performance converts into cost savings and emissions reductions—useful for benefit-cost and investment appraisal of AI projects.
• Importance of organizational frictions in diffusion models
  • Data readiness and workforce AI literacy materially condition returns; models of AI adoption should incorporate these complementary investments (data infrastructure, training) as binding constraints and partial determinants of realized ROI.
• Policy design and industrial policy
  • Public interventions (subsidies, low-cost financing, training programs, digital infrastructure investments) can raise adoption rates and social returns by addressing upfront costs and skill gaps—especially critical in emerging economies where private incentives may underinvest.
• Carbon accounting and externalities
  • Forecast-driven reductions in overproduction, waste, and transport emissions suggest AI investments can be framed (and possibly subsidized) as climate-mitigation measures; economic evaluations should internalize these externalities.
• Future research directions relevant to AI economics
  • Multi-firm, cross-sector replication to estimate heterogeneity in returns and general equilibrium effects.
  • Dynamic adoption models: how learning, trust, and human–AI complementarities evolve and affect long-run productivity and labor outcomes.
  • Cost-benefit analyses that include transition costs (re-skilling, cybersecurity, governance) and distributional effects across workers and suppliers.
  • Incorporate uncertainty and robustness: evaluate performance under structural breaks, supply shocks, and adversarial data conditions.

Summary takeaway: In a data-constrained emerging-market manufacturing setting, a pragmatic hybrid CNN–LSTM deployment produced substantial dual gains (economic + environmental), but these gains depended critically on complementary investments in data integration and workforce capabilities. Cost–benefit and adoption models in AI economics should therefore internalize organizational readiness and environmental externalities to capture the full value of predictive AI.