Simple models beat AI nationally, but machine learning shines where markets are complex: linear regression predicts agricultural credit volatility best overall in Ukraine (2015–2020), yet ANN and gradient boosting improve accuracy by up to 10.6 percentage points in shock-prone regions, supporting a 'precision banking' hybrid approach.

Main Finding

• At the national (aggregated) level for Ukrainian agricultural loan interest-rate forecasting (2015–2020), a simple Linear Regression (LR) baseline outperformed modern AI methods (ANN, Random Forest, Gradient Boosting, SVR, KNN) with average accuracy 82.35% (MAPE 17.65%).
• However, in economically complex, high-volatility regions (e.g., Kherson, Dnipropetrovsk, Zaporizhzhia), AI models—especially ANN and Gradient Boosting—captured nonlinear dynamics missed by LR and improved forecast accuracy by up to 10.6 percentage points.
• Conclusion: The value of AI is context-dependent — parsimony (simple models) is preferable in stable environments, while AI yields clear benefits in volatile/structurally changing markets. The authors propose a “precision banking” hybrid approach (linear models in stable regions; AI in high-volatility regions).

Key Points

• National aggregate results (average accuracy):
  • Linear Regression: 82.35% (baseline)
  • SVR: 74.58%
  • KNN: 74.34%
  • Random Forest: 71.57%
  • Gradient Boosting: 71.45%
  • ANN: 71.37%
• Regional heterogeneity:
  • Stable regions (examples): Luhansk 94.13% (LR), Lviv 92.53% (LR), Chernihiv 92.22% (LR).
  • Volatile/complex regions (examples): Kherson — ANN 92.47% vs LR 85.06%; Zaporizhzhia — ANN 84.09% vs LR 73.49% (10.6 pp gain); Dnipropetrovsk — Gradient Boosting 86.81% vs LR 81.43%.
• Interpretation:
  • Principle of parsimony/Occam’s razor explains LR superiority when the underlying process is a strong deterministic trend (low volatility).
  • AI models are prone to overfitting in low-complexity settings but can detect nonlinearities and structural breaks when data complexity/volatility is high.
  • Results align with the Adaptive Markets Hypothesis: the marginal utility of sophisticated models rises with market volatility and complexity.
• Policy/operational recommendation: adopt a hybrid "precision banking" model—keep simple, interpretable models for stable regions and deploy AI-based risk detection in volatile regions to improve credit allocation and pricing.

Data & Methods

• Data
  • Monthly regional interest rates on loans to firms in Ukraine’s agricultural sector, covering 25 administrative regions plus Kyiv for 2015–2020.
  • Per-region sample: 72 months total; training = Jan 2015–Dec 2019 (60 months), testing = Jan–Dec 2020 (12 months).
  • Features: time index, calendar-year indicators, short lags (Rt−1, Rt−2, Rt−3) and seasonal lag (Rt−12).
• Models compared
  • Baseline: Simple Linear Regression (time index and trend).
  • Machine learning: Random Forest (100 trees), Gradient Boosting Regressor, Support Vector Regression (SVR), K-Nearest Neighbors (KNN).
  • Deep learning: Multi-layer Perceptron (ANN) with two hidden layers (100 and 50 neurons), ReLU activation, Adam optimizer, Z-score normalization; max iter 1000.
• Evaluation
  • Recursive window forecasting on the 2020 test set (forecast each month using available history), mimicking real-time forecasting.
  • Performance metric: Mean Absolute Percentage Error (MAPE); reported as Accuracy (%) = 100 × (1 − MAPE). Test set n = 12 per region.
• Noted methodological limitations (as reported or implied):
  • Small out-of-sample test window (12 months per region).
  • Hyperparameter tuning details sparse; fixed architectures/hyperparameters reported.
  • Dataset ended in 2020 (pre/post major shocks like the 2022 war not covered); generalization to later/post-war periods not directly evaluated.

Implications for AI Economics

• Conditional value of AI: The economic payoff from deploying complex AI models for financial risk forecasting depends strongly on local volatility and structural complexity. Blanket adoption of AI is not cost-effective when underlying processes are stable and largely linear.
• Model selection as a regional policy lever: Financial institutions and regulators should adopt geographically differentiated modeling strategies (precision banking) to improve pricing, reduce exclusion, and limit misallocation of credit. Simpler, interpretable models are adequate (and cheaper) in stable regions; AI should be targeted to volatile regions where it materially improves risk detection.
• Financial inclusion and pricing: Mispricing risk by using a uniform model across regions can lead to over-collateralization, exclusion of smallholders, or unnecessary tightening in fragile regions. Targeted AI application can improve credit access and better align collateral and lending terms with local risk.
• Regulatory and interpretability trade-offs: The paper reinforces the ongoing AI-economics debate: higher accuracy in complex regions may come at the cost of interpretability. Regulators should balance the gains of AI-driven accuracy against the need for transparent, auditable risk models—especially in banking.
• Research and deployment suggestions:
  • Invest in regime-change detection (structural-break detectors, monitoring) to trigger migration from simple to AI models when volatility rises.
  • Combine AI with explainability tools (e.g., SHAP) for deployment in regulated environments.
  • Extend analysis with richer features (macro, weather, input prices, conflict indicators) and longer post-2020 data to assess AI value during larger shocks (e.g., wartime effects).
  • Conduct cost–benefit analyses accounting for model training, maintenance, and interpretability/regulatory compliance costs to guide practical adoption decisions.

Summary takeaway: This paper provides empirical evidence that the economic value of AI in agricultural financial risk forecasting is heterogeneous: simple econometric models can outperform AI on average and in stable regions, while AI delivers significant marginal gains in volatile, complex locales—supporting a hybrid, precision-targeted approach to model deployment.