A Neural Networks Model of the Venezuelan Economy

Besides an indicator of the GDP, the Central Bank of Venezuela generates the so called Monthly Economic Activity General Indicator. The a priori knowledge of this indicator, which represents and sometimes even anticipates the economy’s fluctuations, could be helpful in developing public policies and in investment decision making. The purpose of this study is forecasting the IGAEM through non parametric methods, an approach that has proven effective in a wide variety of problems in economics and finance.

💡 Research Summary

**
The paper presents a forecasting system for Venezuela’s Monthly Economic Activity General Index (IGAEM), a monthly composite indicator published by the Central Bank of Venezuela that serves as a leading proxy for GDP movements. The authors adopt a non‑parametric, neural‑network‑based approach, arguing that such methods have proven effective in a wide range of economic and financial prediction problems.

Data and Sample Splits
The dataset comprises monthly IGAEM observations from January 1991 through December 2003. Because the series is only available for this period, the authors split it into a training window (January 1992 – December 1999, 96 observations) and a testing window (January 2000 – December 2003, 48 observations). The training period is used to fit the neural models, while the testing period evaluates out‑of‑sample predictive performance.

Model Architecture
The forecasting system consists of nine multilayer perceptron (MLP) networks. Eight “sub‑networks” each receive a distinct subset of macro‑economic and commodity variables as inputs; the ninth, called the “Master Network,” takes the eight sub‑network outputs as its inputs and produces the final IGAEM forecast. This hierarchical ensemble design is intended to capture diverse patterns in the data and to improve robustness by aggregating the predictions of several specialized models.

Input Variables
The authors select a broad set of explanatory series that are believed to influence the Venezuelan economy, especially through oil‑price channels. The variables include:

• Power consumption (GWH) – national electricity usage.
• Caracas Stock Index (IBC).
• Monthly loan rate (BCV bulletin).
• Light crude oil price (CL).
• Consumer price index (IPC).
• S&P 500 index.
• 90‑day Treasury bill rate (T‑Bills).
• Gold price (100 oz).
• Copper price.
• Eurodollar rate.
• Commodity Research Bureau (CRB) index.
• Dow Jones Utilities index.

Each sub‑network is fed a different combination of these series, thereby exploring various lag structures and cross‑asset relationships.

Pre‑processing
To mitigate non‑stationarity and noisy fluctuations, the raw series are smoothed using moving averages and block averages. The authors also normalize the data to comparable scales before feeding it to the networks. While they note that back‑propagation is relatively tolerant of imperfect data, the preprocessing steps are presented as essential for facilitating learning.

Training Procedure
The MLPs are trained with standard back‑propagation and gradient descent. Random initial weights are drawn from a symmetric interval, and a learning rate is chosen empirically. The paper briefly mentions the universal approximation theorem (Cybenko, Hornik) to justify the use of a single hidden layer, but it does not disclose the exact number of hidden neurons, activation functions, number of epochs, or any regularization technique. Consequently, the risk of over‑fitting is not fully addressed.

Performance Metrics
Five evaluation criteria are employed:

1. Hit Rate – percentage of periods where the predicted direction (up or down) matches the actual direction. The system achieves >80 % hit rate on the test set.
2. Efficiency – realized profit as a percentage of the theoretical maximum profit if the series were perfectly forecastable.
3. Mean Error (EAM) – average absolute deviation between forecast and actual IGAEM values.
4. Mean Quadratic Error (ECM) – root‑mean‑square deviation, giving more weight to larger errors.
5. Modified Sharpe Ratio (SRM) – efficiency divided by average draw‑down, intended to balance profitability against volatility of losses.

The Master Network consistently outperforms each individual sub‑network on the Sharpe‑type measure, suggesting that the ensemble aggregation yields a more stable and profitable forecasting rule.

Results
The authors report that the Master Network’s hit rate exceeds 80 %, and its Sharpe‑type ratio is “way above” those of the sub‑networks. They also present tables showing that the IGAEM’s directional agreement with GDP improves in more recent sub‑samples, and that mean absolute and quadratic errors decline as the sample becomes more contemporary. These observations are used to argue that the IGAEM is a reliable leading indicator of GDP and that the neural‑network system captures genuine patterns rather than random noise.

Critical Assessment

• Data Limitations – The entire sample contains only 144 monthly observations, which is modest for training multiple neural networks with many inputs. This raises concerns about over‑fitting, especially given the lack of explicit regularization or cross‑validation.
• Model Transparency – Key architectural details (hidden‑layer size, activation functions, training epochs, early‑stopping criteria) are omitted, making replication difficult.
• Benchmarking – No comparison is made with conventional time‑series models (ARIMA, VAR, simple linear regression) or with other machine‑learning approaches (support vector regression, random forests). Consequently, the claimed superiority of the neural ensemble cannot be quantified relative to standard baselines.
• Variable Selection & Lag Structure – The choice of input variables and their lag lengths appears ad‑hoc. No statistical tests (e.g., Granger causality, variance inflation factor) are reported to justify inclusion or to address multicollinearity.
• Evaluation Scope – While hit rate and Sharpe‑type ratios are informative for directional trading strategies, they do not fully capture forecast accuracy for policy‑making purposes. Metrics such as Mean Absolute Percentage Error (MAPE) or out‑of‑sample R² would provide a more complete picture.
• Economic Interpretation – The paper emphasizes the practical relevance of early IGAEM forecasts for policy and investment, yet it does not discuss how the model’s predictions could be integrated into concrete decision‑making frameworks (e.g., fiscal rule adjustments, portfolio allocation).

Future Directions

1. Extended Sample – Incorporate more recent IGAEM data (post‑2003) and possibly higher‑frequency indicators to increase the training set.
2. Robust Validation – Apply k‑fold cross‑validation or rolling‑window out‑of‑sample tests to assess stability over time.
3. Regularization Techniques – Use dropout, L1/L2 penalties, or Bayesian priors to mitigate over‑fitting.
4. Benchmark Comparisons – Evaluate the neural ensemble against ARIMA, VAR, and modern machine‑learning models to establish relative performance.
5. Feature Engineering – Conduct systematic lag‑selection, principal component analysis, or mutual information screening to reduce dimensionality and multicollinearity.
6. Economic Integration – Translate forecast outputs into actionable policy signals (e.g., early‑warning thresholds for fiscal deficits) and assess the economic value of such signals through scenario analysis.

Conclusion
The study demonstrates that a hierarchical ensemble of neural networks can achieve high directional accuracy (over 80 % hit rate) in forecasting Venezuela’s IGAEM, a leading indicator of GDP. The Master Network’s superior Sharpe‑type performance suggests that aggregating diverse sub‑models yields a more reliable predictor. However, methodological gaps—particularly regarding data sufficiency, model transparency, regularization, and benchmark comparison—limit the confidence with which the results can be generalized. Addressing these issues in future work would strengthen the case for neural‑network‑based macro‑forecasting in oil‑dependent economies like Venezuela.