Falling for risk management
Two papers on drawdown geometry and sovereign wealth funds shifted my view of risk management from peripheral to central.
Risk management used to feel peripheral to me. Drawdown limits and value-at-risk reports read like compliance furniture. The mathematically interesting work seemed to live on the alpha side. Two recent projects changed that view.
The first was REBOUND. Most allocators minimize variance, but variance does not capture the shape of a loss: how deep it goes, how long it lasts, whether it recovers. Treating drawdown geometry as a first-class objective changes the optimisation problem. Instead of minimising spread around a mean, the question becomes what kind of recovery dynamics the portfolio can underwrite. Across 31 years of multi-asset data, the resulting allocator achieves Sharpe 1.54 and max drawdown 9.11%, compared to 0.52 and 55.19% for buy-and-hold.
The second is the sovereign wealth funds paper, currently under review at the Journal of Economic Surveys. The literature largely treats SWFs as static allocation pools. The interesting question is what they do during crises: whether they transmit volatility into the host economy or absorb it, whether their drawdowns force liquidations elsewhere. That is a control problem under regime uncertainty. The failure modes are not bad returns but stability events in real economies.
Both projects converge on the same point. The questions worth taking seriously in finance are not whether a model outperforms but whether it survives. Survival is a strictly different objective. It requires taking regimes seriously instead of averaging over them, taking tails seriously instead of trimming them, and treating the cost of being wrong as part of the optimisation rather than a noise term.
This is not a move away from machine learning. Reinforcement learning remains the most natural framework I know for sequential decisions. But the version of RL I now find most interesting is the risk-aware one: policies that satisfy survival constraints, reward shaping that respects transaction costs, and inverse RL that recovers what the demonstrator could not afford to lose. RRR is built on that view. The journal extension of ARISE shares it.
I am pursuing PhD positions and quant or risk-management roles for the 2026 cycle. If you work on these problems, I would like to hear about it.