Warning - it can be tough reading (and not in the usual scholarly sense, either!) as the author definitely had his own "style". Also it's written from a Canadian point of view, though the concepts are universal. (ie - Americans shouldn't automatically dismiss the link.)

The critique is aimed at the assumption that your risk aversion is scale invariant. i.e. you behave the same when the values are in the 10s of dollars, or the 10,000s of dollars. I might be perfectly fine with taking the coin flip when the outcomes are either $10 or $15, but if the outcomes are $10,000 or $15,000 I might rather take a lower guaranteed amount of $12,000 because that will meet my expenses but the $10,000 won't.

- suppose you are risk neutral (γ = 0). You just maximize cash flow. The solution the model will output (modulo numerical noise): invest 100% in equities; spend 0 each period until last period, when you spend 100%.

- maybe you truly just want to maximize expected value, and that's your solution, and you don't need a complex model.

- But maybe you value spending cash smoothly over time. Now you have a tradeoff: you can spend more smoothly, and spend some cash early, but that will reduce your overall cash flow. How do you decide how to trade them off?

- that's the role of γ. I would just view it as a tunable parameter that makes a rational tradeoff between spending as much as possible, and as smoothly as possible.

- as you increase γ, the model gradually increases bonds more and does so earlier during retirement, reduces variable spending, and increases constant spending to smooth cash flow.

- as γ → ∞, you are left with perfectly smooth cash flow, and an allocation that maximizes the amount you could have spent without ever running out of cash during the historical period - something like the Bengen 4% rule.

The genesis was that I looked at the literature, saw there were a lot of relatively ad-hoc studies of arbitrarily rules, and asked 'what would Google do', and the answer is to first decide what is your cost function, what are you trying to maximize or trade off.

This is one answer...there are more complex answers, one could use a life table and maximize over all mortality scenarios, one could penalize any decline in spending per a prospect theory utility function, etc., etc.

Main takeaway is, with tools like TensorFlow, you can optimize even pretty complex path-dependent financial questions, which you really couldn't do even a couple of years ago, at least not without some really fancy parallel hardware and programming.

It's a first real attempt at a real TensorFlow model, might not conform 100% to best practice, any pull requests appreciated.