This isn't quite right.
Kelly was ostensibly the 1950s-era brainchild of one John L. Kelly of Bell Labs.
In reality, however, the theoretical underpinnings of Kelly had already been developed by Bernoulli in the 18th century (whose methodology of the maximization of geometric mean was ultimately quite similar to Kelly’s) and then later formalized within the context of the isoelastic utility and constant elasticity of substitution production within microeconomics (I’d have to really hunt around to figure out it its true originator – Keynes, perhaps?)
Although Kelly most likely developed his criterion without prior knowledge of the existing work on the subject, it's nevertheless only original in the sense that he explicitly applied it to gambling (and
only gambling -- horse betting in particular).
Kelly's derivation (and you can download his original paper in PDF format
here) while extremely clever, is clearly not the work of an economist. Kelly doesn't derive a solution by appealing to the preferences of a gambler but rather by assuming that maximizing expected bankroll growth is a self-evident logical goal in and of of itself for
any bettor able to reinvest profits. That such a strategy might not be applicable for certain classes of risk averse bettor never enters in to Kelly’s discussion.
As such Kelly doesn't touch on “fractional Kelly” (which doesn't have quite so intuitive an explanation as simply "maximizing expected bankroll growth") and instead talks only about growing generic "capital" or "wealth" exponentially. The precise meanings of these two phrases are never explicitly defined and I suppose are just meant to be taken axiomatically. Nevertheless, by examining Kelly from an economic perspective we can get a clearer picture.
The n-Kelly functions are the sole members of a class of a set of utility functions known as
isoelastic utility. These are functions depending on a single parameter α and are of the form:
U(x;α) = α/(α-1) * x (α-1)/α for α ≠ 1, > 0
U(x;α) = ln(x) for α = 1
so that the marginal utility (i.e., the first derivative of utility with respect to x, corresponding to the increase in utility of obtaining slightly more of x) will equal:
Ux = x-1/α for α > 0
One feature of this class of function is that they exhibit what's known as "constant relative risk aversion". This means that regardless of any increase or decrease in bankroll a player's aversion to risk (his dislike for outcome uncertainty) will remain unchanged when evaluating a given percentage change in bankroll (constant
absolute risk aversion, on the other hand would correspond to equivalent aversion to risk for a given
dollar change).
Putting it another way, if two individuals both exhibit isoelastic utility, each with the same α parameter, then each will treat an x% change in bankroll identically irrespective of their relative bankroll levels. In other words, if I have $10,000 to my name and you have a $1,000,000 then an additional $100 would increase my happiness by the same amount as an additional $10,000 would increase yours.
Now whether you agree with the above depiction or not is largely irrelevant. These are the utility functions implied by Kelly (with α representing the so-called "Kelly fraction") and if this depiction seems inapplicable to your personal preferences, then while that may well be a trait that you share with most human beings, it in no way implies the inefficacy of n-Kelly at maximizing the utility of a hypothetical individual whose preferences
are in line with isoelastic utility.
As I alluded to earlier (and have mentioned countless times on this forum), one by-product of full-Kelly maximization is that it will also maximize the expected growth rate of a (still axiomatic) bankroll. This means that over an arbitrarily long stretch of time the probability that bankroll being bet with Kelly will be larger than a bankroll bet using some other strategy will approach certainty (100%)..
While this may be a by-product of isoelastic utility, it was nevertheless the precise and only effect intended by Kelly. Kelly didn't consider any of the other implications of this mode of thought, nor did he even consider its applicability. He simply made the a priori declaration that such a strategy was effectively preference-independent (I mean who wouldn't want to maximize bankroll growth? Duh.) Now that in no way is meant to marginalize the contributions of Dr. Kelly. The guy wasn't an economist --he just derived his results from the perspective of maximizing expected growth, rather than from the perspective of maximizing a broader class of utility function.
So really there are two schools of thought from which to consider Kelly:
- the John L. Kelly/Bernoulli manner -- Kelly is the strategy maximizing the expected growth of a bankroll. Fractional Kelly, while not explicitly defined, can be considered the full-Kelly stake multiplied by the Kelly fraction. The fractional-Kelly stake may not be well-defined for all classes of bets, nor may all questions of Klely-optimality be easily answered.
- the utility function manner -- Kelly is the strategy maximizing the utility function for a bettor with Kelly risk aversion parameter α. The fractional Kelly stake will always be well-defined for any class of bets, and in general, for small stakes, will closely mirror the fractional Kelly stake defined in 1) above. This methodology is readily extensible to include all forms of wealth manipulation.
So what I’m going to argue is that neither of these interpretations is consistent (in most cases) with an easily replenished bankroll.
In the Kelly/Bernoulli school the Ultimate Goal of Existence is getting as rich as possible as deliberately and as quickly as possible. Losing one's entire is infinitely bad (log(0) ~ -∞) because once one's bankroll drops to 0 it's impossible for it to grow ever again. It's not just bad, or very bad, or really bad, really really very very bad, but
infinitely, no possible to do any worse than this, Book of Revelations bad. Now that's bad.
Even were a bettor's wealth to drop to only $0.01, then as long as she could continue making bets (i.e., there are no betting minimums) then she still has a possibility of growing it again (and in fact will do so with probability approaching 1 over an increasing time horizon). So while this may be "bad" it's still
infinitely better than losing that last penny from which no recovery is possible.
Within this interpretation it's not about being happy or sad but about the ability to grow one's bankroll. If however, another source of betting capital existed (for example, you just needed to transfer more money from your checking account) then losing that penny is necessarily stripped of its "God Almighty Has Struck You Down" bad-ness. While going broke in this limited sense might "suck" (or even "suck really, really hard") it does
not represent a situation from which no recovery is possible and as such can not correspond to a negative Kelly-score.
In the microeconomics Kelly-as-isoelastic-utility School the Goal of Existence is to maximize a utility function for no existential purpose other than "that's what people do". Bankroll growth is a necessary and sufficient by-product of full-Kelly maximization (like love and marriage) but does not follow from fractional-Kelly maximization (for fractions other than 1).
Losing one's entire bankroll is again infinitely bad (for Kelly mulitiples ≤ 1) because marginal utility grows unboundedly larger as one's bankroll approaches zero. As such, a Kelly bettor will never risk
any possibility of going broke (regardless of current wealth) and such an outcome would result in "sadness unyielding".
What's must illustrative, however, is the limiting case where win probability of a given bet (holding payout odds constant) approaches 100%. What we see is that as the probability of winning a bet increases (and we're not just talking from 90% to 95%, but more like from 99.9999999% to 99.99999999%) a bettor will become increasingly more willing to bet every dollar on which he can get his hands. Borrow from the in-laws, borrow from your in-laws’ in-laws, raid your daughter's college fund, second and third mortgage your house, second and third mortgage your granddaughter’s virginity, etc.)
This provides the clearest indication of what a "complete Kelly-bankroll" really is. If the consequences of losing sound disastrous well that's precisely the point. As I mentioned before, losing a bankroll can't just suck, it has to
suck infinitely. It has to suck so bad that the only time that a bettor would ever be willing to risk all of it would be if he possessed
100% certainty that he'd win.
Now I fully grant that this interpretation might not be computationally convenient. But that’s beside the point. Convenience and intellectual probity, after all, don’t necessarily go hand-in-hand. From an operational perspective what should be taken from this is that a Kelly bankroll should be maximally construed (and remember that this doesn't take into account withdrawals from one's bankroll for the purposes of living expenses) and the real fudging should come in the form of adjusting one's Kelly multiplier.
If you're (truly) a professional bettor with a $1,000,000 betting bankroll and $3,000,000 in savings you're probably better off assuming a $4MM bankroll with around a 5% Kelly fraction than a $1MM bankroll with around a 20% Kelly fraction (and in this simplified analysis we're admittedly neglecting the costs of transferring money from savings to betting accounts, as well as the opportunity costs (in terms of interest, dividends, capital appreciation, etc.) of removing capital from savings). Most of the time the difference between the two will be minimal, however where the contrast may well become apparent would be in times of several exceptionally good bets or during a period where a huge number of fairly good bets come along (such as, perhaps, during the Super Bowl or March Madness).
As I've said before this is certainly a rather academic point, but that doesn't detract from its representing an unsullied view of the Kelly criterion. While any given individual’s personal preferences may well deviate
substantially from those implied by Kelly, that in no way implies those preference more indicative of Kelly than those I’ve outline above.
I will grant, however, that there are some circumstances where within Kelly "wealth" may be narrowly construed. These situation will follow from scenarios such as those to which you allude where a bankroll is taken in a vacuum. While such a scenario could fit in with the Kelly-Bernoulli interpretation where one's trying to grow an isolated set of capital for no particular reason (if it were general wealth enhancement which one were after, then isolating the bankroll would be suboptimal) it's difficult to see how it could fit in with general concepts of utility maximization in all but the most of contrived of examples.
All I can think of would involve two individuals competing in a bankroll growth contest for a fixed prize given some arbitrary starting values. Obviously, in such a contest considering any other wealth not available to the contest itself would be meaningless.
The isoelastic utility view of Kelly is simply an extension of the very well-defined preference-theory of financial economics. And the only straightforward way for the conclusions of Kelly to match the implications stated isoelastic preferences is to construe bankroll/wealth/capital in a certain manner. This isn't necessarily most convenient (or least dangerous for a self-deluding "advantage" bettor) but it is nevertheless most accurate and in keeping with underlying theory.
While all these criticisms of Kelly might be difficult to refute utilizing the narrow Kelly/Bernoulli interpretation, they simply become issues of increasing computational complexity when the Kelly criterion is considered as isoelastic utility maximization. I believe I’ve touched on all these concepts on this forum.