💻 computer science

A TimeGAN Framework for Synthetic Frontier-Market Currency Scenarios: Evidence and Critical Reflections from the Nigerian Naira/US Dollar Market

This paper critically evaluates a TimeGAN framework for generating synthetic Nigerian Naira/US Dollar scenarios under data scarcity, revealing that the model's outputs are dominated by classical baselines in distributional fidelity and exhibit extreme seed-dependent instability, thereby arguing that GAN-based scenario generation requires specific architectural priors and rigorous multi-seed reporting before it can responsibly complement conventional stress-testing tools.

Original authors: Tosin Bello

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tosin Bello

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of finance, risk managers act as the guardians of stability, constantly trying to predict how much money could be lost if a market suddenly turns against them. To do this, they rely on history, looking at past price movements to guess what might happen tomorrow. However, this approach hits a wall when dealing with frontier markets, such as the Nigerian naira. These economies often have very short, patchy records of data, and their currencies can swing wildly due to sudden political or economic shifts. When there is not enough history to learn from, analysts sometimes turn to artificial intelligence to create synthetic data—computer-generated scenarios that mimic real market behavior. The hope is that these machines can fill in the gaps, creating enough plausible "what-if" stories to test how a portfolio would survive a crisis. One popular tool for this is a type of AI called a Time-series Generative Adversarial Network, or TimeGAN. It works by having two neural networks play a game against each other: one tries to create fake market data, and the other tries to spot the fakes. Over time, the creator gets so good that the data looks indistinguishable from the real thing, theoretically allowing experts to stress-test their investments against thousands of imagined futures.

A researcher at the University of Ibadan set out to test whether this technology could actually work for the Nigerian naira, a currency that has faced severe turbulence in recent years. The goal was to build a system that could generate realistic scenarios for the naira's exchange rate against the US dollar, using only the limited, publicly available data an independent researcher could legally access. The study focused on a specific period from June to September 2026, a time when the naira had settled into a relatively calm phase after a major devaluation crisis. The researcher trained the AI on just ninety days of daily price changes, a tiny amount of information compared to the thousands of data points usually required for such models to work well. To see if the AI was doing its job, the results were compared against a more developed market currency, the euro against the dollar, and against traditional, well-established statistical methods that have been used for decades.

The investigation began with a routine check that changed the entire direction of the study. When the researcher ran the same computer code twice on the same data, expecting the exact same result, the numbers came out different. One run suggested the currency was safer than it actually was, while the next run suggested it was much riskier. This inconsistency revealed a fundamental problem: the AI was not learning a single, stable truth from the data. Instead, its output was heavily influenced by the random starting point, or "seed," used to initialize the computer's calculations. Realizing that a single result could be a fluke, the researcher abandoned the standard practice of reporting just one outcome. Instead, the study was redesigned to run the model five times with different random seeds for every single number reported, treating the variation between these runs as a critical part of the finding.

When the results were finally analyzed using this stricter, multi-run approach, the AI failed to outperform the traditional methods. The synthetic data it produced did not match the shape of the real market history as well as the older, simpler statistical models did. In fact, a computer program trained to tell the difference between real and fake data could spot the AI's fakes with perfect accuracy every single time. The AI consistently failed to capture the extreme, rare events that are most important for risk management, creating data that was too smooth and too predictable compared to the messy reality of the market. Most importantly, the risk estimates produced by the AI were wildly unstable. Depending entirely on which random seed was used, the model's prediction for the potential loss of the naira varied by a factor of roughly two and a half. One run might suggest a loss of just over one percent, while another run on the exact same data suggested a loss of nearly three percent. This range was so wide that it covered the predictions of every other method tested, from the traditional models to simple historical averages.

The study concludes that for a frontier market with scarce data, this type of advanced AI is not yet ready to be used as a standalone tool for safety checks. The lack of historical information leaves the model "under-identified," meaning there is not enough evidence in the data for the computer to settle on a single, reliable answer. Instead of producing a clear picture of the future, the model produces a fog of possibilities where the outcome changes based on a random number the researcher cannot control. The researcher argues that before such technology can be trusted in a real-world risk office, it needs to be combined with other methods, trained on much larger datasets from more stable markets, and always tested across many different random starting points to ensure the results are not just luck. Until then, the most reliable way to understand the risk of a currency like the naira remains the traditional statistical approaches, which, while simpler, offer a stability that the current generation of AI cannot yet provide.

Technical Summary: A TimeGAN Framework for Synthetic Frontier-Market Currency Scenarios

Problem Statement
Frontier-market currencies, such as the Nigerian Naira (NGN), face a critical methodological challenge: they are highly exposed to sudden depreciation and regime shifts, yet their historical data is typically short, fragmented, and inaccessible. While Generative Adversarial Networks (GANs), specifically Time-series GANs (TimeGAN), have been proposed as data-driven alternatives to classical stochastic models (e.g., GARCH, Black-Scholes) for generating synthetic financial scenarios, their efficacy in data-scarce frontier markets remains unverified. Most existing literature relies on multi-year, high-frequency datasets from deep, liquid markets (e.g., S&P 500, EUR/USD). This paper investigates whether a TimeGAN framework can reliably generate synthetic NGN/USD scenarios under the strict data constraints faced by independent researchers, specifically testing if the model can outperform classical baselines when trained on limited, non-stationary data.

Methodology
The study constructs a compact TimeGAN pipeline and benchmarks it against classical stochastic models using a rigorous, multi-seed reproducibility protocol.

Data Sources:
- Target: Daily NGN/USD mid-market rates from CurrencyBeacon (June 18 – September 16, 2026), yielding 90 daily log returns. This period represents a post-unification stabilization phase, not the 2023–2024 crisis.
- Benchmark: Daily EUR/USD rates from the Federal Reserve H.10 series (January 2018 – March 2020), yielding 555 returns, which includes the March 2020 COVID-19 volatility shock.
- Constraints: The NGN series is short (90 observations) compared to the thousands typically used in GAN literature. No GPU acceleration was used; training occurred on a single CPU core to reflect independent researcher resource profiles.
Model Architecture:
- A compact TimeGAN (Yoon et al., 2019) comprising five networks: an Embedder, Recovery, Generator, Supervisor, and Discriminator.
- All networks utilize single-layer Gated Recurrent Units (GRUs) with dense output layers.
- Hidden dimensions were set to 6 for NGN/USD and 10 for EUR/USD.
- Training followed the three canonical phases: autoencoder pretraining (reconstruction loss), supervised pretraining (next-step prediction in latent space), and joint adversarial training.
Experimental Protocol (The Five-Seed Pivot):
- A critical methodological shift occurred during the study. An initial single-run verification revealed that the model's output varied significantly based on random initialization.
- Consequently, the protocol was changed to train five independent models per series using distinct random seeds (101, 202, 303, 404, 505).
- Results are reported as the mean, standard deviation, and range across these five seeds, rather than a single point estimate. Classical baselines (GARCH(1,1), Monte Carlo, Historical Bootstrap) are reported as single deterministic values.
Evaluation Metrics:
- Distributional Fidelity: Earth Mover's Distance (EMD), Kolmogorov-Smirnov (KS) test, and moments (skewness, kurtosis).
- Temporal Dependence: Autocorrelation of squared returns (volatility clustering) and leverage effects.
- Discriminative Score: A Random Forest classifier trained to distinguish real vs. synthetic windows (0.5 = indistinguishable; 1.0 = perfect separation).
- Risk Metrics: Value-at-Risk (VaR) and Expected Shortfall (ES) at 95% and 99% confidence levels over the training horizon (10 days for NGN, 20 days for EUR).

Key Results
The study finds that the TimeGAN fails to dominate classical baselines and exhibits fundamental instability in low-data regimes.

Distributional Failure: Across all five seeds for both series, the TimeGAN produced synthetic distributions that were platykurtic (excess kurtosis $\approx$ -1.15 for NGN, -1.03 for EUR), failing to reproduce the fat tails of the historical data (which was mesokurtic for NGN and leptokurtic for EUR).
Discriminative Separability: A Random Forest classifier achieved 100% accuracy in distinguishing real from synthetic windows in all ten training runs (5 seeds $\times$ 2 series). The generator never successfully fooled the classifier.
Volatility Clustering: While the TimeGAN captured short-term clustering (lag 1), it failed to reproduce the multi-week volatility persistence observed in the historical data (lags 2–15), where synthetic autocorrelation collapsed to zero.
Seed-Induced Instability (Central Finding): The most significant finding is the extreme variance in risk estimates based solely on the random seed.
- NGN/USD: The 95% VaR estimates across the five seeds ranged from 1.07% to 2.54% (a factor of 2.4). This range straddled the GARCH estimate (1.06%) and the Historical Bootstrap (1.59%).
- EUR/USD: The 95% VaR ranged from 1.98% to 4.90% (a factor of 2.5), again straddling all classical baselines.
- This implies that two analysts training the identical model on identical data could derive risk estimates differing by more than double, purely due to initialization.

Significance and Claims
The paper does not claim that TimeGAN is a superior tool for frontier markets; rather, it argues that data scarcity renders the model under-identified at current sample sizes.

Reproducibility as a Discovery: The paper claims that the routine reproducibility check (the five-seed protocol) was the primary driver of its findings. Had the study reported a single run, it would have presented a misleading, potentially biased point estimate. The instability itself is the core result.
Regulatory Implications: The findings support the Basel Committee's shift from VaR to Expected Shortfall (ES). Because the TimeGAN produces unstable tail estimates, relying on a single VaR number is dangerous. Reporting ES alongside VaR, and evaluating across multiple seeds, provides a more robust, albeit still imperfect, view of tail risk.
Future Directions: The paper suggests that for GANs to be viable in frontier markets, they require:
1. Transfer Learning: Pretraining on developed-market corpora to impose structural priors.
2. Structural Priors: Incorporating signature-based or graph-based architectures (e.g., Sig-Graph GAN) to encode path regularities.
3. Tail-Aware Objectives: Loss functions explicitly targeting kurtosis and tail behavior.
4. Standardized Reporting: The field must adopt multi-seed reporting to characterize model variance, not just accuracy.

Conclusion
The paper concludes that for a frontier-market risk function, GAN-based scenario generation is not yet ready as a stand-alone tool. It should not be deployed based on a single training run. Instead, it can only be considered as one input among several, triangulated against classical baselines and characterized across many seeds, before its tail-risk outputs can be trusted. The primary contribution is a critical, self-reflective demonstration that without sufficient data or structural constraints, generative models in frontier markets produce unstable, non-reproducible risk estimates.

Technical Summary: A TimeGAN Framework for Synthetic Frontier-Market Currency Scenarios

More like this