embeds feedback between state evolution and control.

conditions decisions on realised states. 

may respond to interim shocks.

is always superior in every possible setting.

continuous validation, refinement, and governance after deployment.

careful calibration of stochastic drivers and scenarios. 

purely increasing network depth and width.

precise specification of the business model and accounting logic. 

It relies on parameters that are difficult or impossible to estimate reliably in reality.

It assumes precise knowledge of expected returns and volatilities.

It presumes frictionless markets without transaction costs or funding constraints.

It adapts dynamically to realised states using observable information only.

critical appraisal of learned behaviour and limitations. 

independent validation on out-of-sample scenarios.

 stress testing under adverse paths. 

blind reliance on training performance metrics.

Poor scenario design can bias the learned strategy.

They define the economic environment in which decisions are trained.

They determine how well the model generalises to stress situations.

They replace the need for any structural modelling assumptions.

it approximates policies instead of value functions on full grids.

it focuses on relevant trajectories rather than exhaustive state-space discretisation. 

 it trades global optimality for tractable approximations.

it eliminates uncertainty entirely.

risk aversion is explicitly parameterised.

 the optimal policy depends on the chosen risk parameter.

only expected return matters. 

the dispersion of outcomes affects optimal decisions.

exponential growth of runtime with problem dimension. 

linear growth of runtime with more and more variables.

the limited number of variables that Deep ALM can account for. 

a limitation of classical dynamic programming methods.