Predicting the bankruptcy of a country earlier makes it possible to limit financial losses, to better prepare political and economic responses and, above all, to mitigate the social consequences. For international institutions and governments, this is an opportunity to put in place aid or restructuring plans before the crisis explodes. And for the population, it can avoid brutal measures: banking collapse, cuts in public services, sudden rise in unemployment or soaring prices. In short, anticipating means gaining time to act with greater control, reducing economic and social damage, and avoiding managing the crisis in a hurry and panic.
In March 2020, as markets collapsed in the midst of a pandemic, few institutions anticipated the scale of the crisis for emerging countries. However, in the following months, Zambia, Sri Lanka and Lebanon found themselves in default of payments. Today, with the rise of artificial intelligence (AI), a question arises: can AI predict state bankruptcy better than traditional models?
The limits of traditional models
So far, major warning tools – such as the early warning system (early warning system) of the International Monetary Fund (IMF) – were based on classic macroeconomic aggregates: external debt, growth, foreign exchange reserves, current account balance, etc. Although useful, these indicators are often published late and susceptible to accounting manipulation.
Rating agencies (S&P, Moody’s, Fitch, etc.), for their part, still largely base their evaluations on human analyses, with a significant time lag.
An inability to pick up on behavioral cues
In fact, traditional models for predicting sovereign risks are mainly based on aggregate macroeconomic indicators: debt/GDP ratio, level of foreign exchange reserves, current account, inflation or even credit agency ratings. These approaches, often inspired by classic econometric or statistical models (logit, probit, etc.), have two major limitations:
- Data too slow and too aggregated. Macroeconomic statistics are published with delays sometimes of several months, quarters, or even years. They smooth out weak signals and mask short-term dynamics such as massive capital withdrawals or emerging bank panics.
- An inability to pick up on behavioral and policy signals. Sovereign crises are not only economic. They are also social, political and sometimes geopolitical. However, traditional models struggle to integrate non-quantitative variables such as political instability, social polarization, protest movements or discreet negotiations with international donors.
The fact that Lebanon was rated B until 2019, when it probably deserved a D, illustrates several major dysfunctions in traditional sovereign rating systems. Agencies prefer to react rather than prevent, which biases their ratings upwards, especially for fragile countries.
Weak signals detected
New models are emerging, based on machine learning (machine learning) and natural language processing (NLP). The rating agencies Fitch, Moody’s and S&P are already testing AI capable of processing thousands of sources of information in real time: financial flows, public declarations, but also satellite data, anonymized Swift transactions and comments on the X network (formerly Twitter).
These tools detect weak signals, invisible to the eyes of traditional analysts: a series of transfers to offshore accounts, an abnormal drop in banking volumes or even a sudden rise in hashtags such as #default or #bankrun in a local language.
The Lebanese precedent
In 2019, well before the official collapse of the Lebanese banking system, several weak signals were visible. Rumors were circulating on WhatsApp and Twitter. Outgoing remittances were exploding. Videos showed protesters demanding accountability from the Central Bank. However, the rating agencies were slow to warn, as was the International Monetary Fund.
Despite alarming signs upstream, the IMF remained cautious in its language and did not issue a frank warning before the collapse of the Lebanese banking system in 2019-2020. An AI-based model trained on non-linear cash withdrawal behaviors, mentions of financial panic on social media, or the gradual disappearance of imported products could have detected a high risk of default well before the agencies.
The case of Pakistan: real-time surveillance
Pakistan, regularly on the verge of payment default, illustrates another use of AI. The laboratory of the Higher School of Foreign Trade (ESCE) has developed a tool combining anonymized banking data flows (deposit volumes by region), satellite data on port activity levels, semantic analysis of political speeches (frequency of terms like “emergency aid”, “moratorium”, “IMF negotiation”), volume of direct outgoing flights from Karachi’s Jinnah International Airport, discussions in Punjabi and Urdu on X and Facebook, etc. By combining these elements, the model anticipated a new request for a rescue plan from the end of 2022… announced publicly only in April.
The asymmetry of power between States and other economic actors maintains the recurrence of systemic financial crises, the impacts of which on major macroeconomic balances are often massive. AI adds a dynamic layer to country risk models, with alerts based on behavioral trends and “low noise.” Where traditional systems struggle to integrate collective psychology or contagion effects on markets, artificial intelligence excels.
AI, a new risk?
Be careful, however, of algorithmic bias. Data from social networks is noisy and its analysis can be influenced by coordinated campaigns. Furthermore, AIs are not neutral: they can integrate unequal representations of country risk, in particular by overestimating tensions in politically or socially unstable, but solvent, countries.
These tools will not soon replace critical eyes and human discernment. The real breakthrough will therefore come from a hybrid model.
Artificial intelligence is not an oracle. Unstructured data (social networks, media, forums) is called “noisy”, in the sense that poor training quality and poor filtering can lead to alarming “false positives”. If the model is trained on past crises, it risks overweighting certain types of signals and underestimating new configurations, or even “hallucinating”. These hallucinations often arise due to problems with the data used to train language models, limitations in model architecture, and the way large language models interpret economic and financial data.
Governance problem
Additionally, sources may themselves be biased by disinformation campaigns. The most efficient AI models are “black boxes”, whose internal logic is difficult to interpret. For an institutional investor or a credit analyst, this can pose a governance problem: how to justify an investment or divestment decision based on an unexplainable algorithmic score?
Faced with the acceleration and increasing complexity of economic, social and geopolitical shocks, institutions such as the IMF, the Fed or the European Central Bank (ECB), as well as certain investment funds and certain banks, are now experimenting with artificial intelligence capable of simulating crisis scenarios in real time. The objective is no longer just to predict the next failure, but to prevent it by adjusting economic policies in a much more responsive way, thanks to adaptive models fed by continuous and dynamic data flows.
But these tools will not soon replace critical eyes and human discernment. Because where AI excels in detecting weak signals and massive data processing, it remains vulnerable to bias and statistical hallucinations if left to its own devices. The real breakthrough will therefore come from a hybrid model: a permanent dialogue between the geopolitical and social intuition of human analysts and the calculation capacity of artificial intelligence.
Ultimately, it is no longer a question of choosing between humans and machines, but of intelligently orchestrating their respective forces. The challenge is not only to see crises coming, but to have the means to act before they erupt.
