For an insurer or a pension fund, investment management is the other half of the actuarial craft. While one side of the balance sheet projects liabilities that mature over decades, the other must build an asset portfolio capable of backing them. Artificial intelligence arrived at that frontier with ambitious promises: forecasting markets, optimizing portfolios, hedging risks dynamically. It pays to separate what AI genuinely does from what it simply wraps in sophisticated language, because in investments a mistake is not paid for in a validation metric but in the institution's capital and in the promises made to policyholders.
The hardest problem: forecasting returns
Let us start where AI disappoints most often. Forecasting the future return of an asset is, at heart, a problem of a very weak signal buried in enormous noise, and a nonstationary one at that: the relationship between variables changes precisely when market participants learn to exploit it. A machine learning model trained on historical prices can show spectacular in-sample fits and collapse out of sample. The fundamental law of active management sums up why:
$$ IR \approx IC \cdot \sqrt{N} $$
The information ratio depends on the information coefficient, that is, the correlation between the forecast and the outcome, and on breadth, the number of truly independent bets. The information coefficient of almost any market signal is tiny, and no neural network architecture changes that arithmetic: at best it exploits breadth that already existed. The value investor knows this by intuition and the actuary by statistical discipline. Distrusting a backtest that looks too good is not pessimism; it is prudence.
Where AI does add value
The fertile ground is not guessing the market but processing information and governing complex portfolios. That is where machine learning delivers concrete, measurable advantages:
- Alternative and unstructured data: reading financial reports, news, satellite imagery or environmental and social metrics at a scale impossible by hand, to enrich credit and issuer analysis.
- Economic scenario generators: calibrating and speeding up the models that feed liability valuation and stress testing, while preserving consistency among interest rates, inflation and credit spreads.
- Risk monitoring: detecting regime shifts, hidden concentrations and correlations that spike precisely when markets come under stress.
- Dynamic hedging: approximating hedging strategies for portfolios with embedded optionality, where no closed-form analytical solution exists.
In all of these cases AI does not replace the investment decision: it broadens the evidence on which the decision is made and frees up the analyst's time for the questions that really matter.
Reinforcement learning and asset liability management
The most promising use case for an actuarial institution is not the hedge fund but asset liability management (ALM). Reinforcement learning frames investing as a sequence of decisions: in each period the manager chooses an allocation, observes how assets and liabilities evolve, and seeks to maximize a long-term reward, for example risk-adjusted surplus. It is no coincidence that the Society of Actuaries (SOA), through its investment section, has published research on deep learning applied to liability-driven investing, and there is already an academic literature on deep reinforcement learning in ALM. Classical immunization requires matching the sensitivity of assets and liabilities to a movement in interest rates:
$$ \frac{\partial A}{\partial y} = \frac{\partial L}{\partial y} $$
Reinforcement learning generalizes that condition to a world with transaction costs, regulatory constraints and multiple objectives, where there is no longer a closed-form formula. The promise is real, but so are the risks: an agent trained inside a simulator inherits all of that simulator's biases and can learn to exploit artifacts of the model instead of managing the true risk.
The discipline that sets the actuary apart
What turns these tools into reliable instruments is not the sophistication of the algorithm but the discipline surrounding it. Rigorously separating training, validation and test data. Penalizing the number of configurations tried before celebrating a result, which is the concern behind the deflated Sharpe ratio. Requiring every signal to have an economic hypothesis behind it, not just a lucky correlation. Watching for data drift when the regime changes. These are principles the actuary already applies to pricing and reserving, now carried over to the investment side of the balance sheet. Model governance, with independent validation, documentation and ongoing monitoring, is as relevant to the portfolio as it is to the liabilities. The Institute and Faculty of Actuaries (IFoA) and the Casualty Actuarial Society (CAS) have developed material on data science and the responsible use of models that frames this responsibility precisely.
Judgment over computing power
When a tool is powerful, the temptation is to let it decide on its own. In investments that delegation is dangerous: the market is not a stationary environment like a board game, and a strategy that worked for a decade can fail just when the most capital is riding on it. The role of the actuary, and of the investor with sound judgment, is not to compete with the machine on calculation speed but to frame the problem, challenge the simulator's assumptions, demand an economic thesis and take responsibility for the outcome before the board and the regulator. Artificial intelligence vastly expands what we can analyze; it does not replace the judgment about what is worth analyzing. In managing the assets that back promises spanning several decades, that difference is everything.