What convex means, without notation
A payoff is convex when improving your outcome by one step is worth more at the top than in the middle. The clean way to see it is to lay the rewards out and look at the gaps. Suppose twelve people compete, most finishes pay nothing at all, and the top finish pays more than the next two combined. Moving from last to eighth is worth zero. Eighth to fourth, zero. Third to second is worth something. Second to first is worth more than everything else on the table.
That is the shape: flat along most of its range, then steep at the very end — the opposite of a payoff proportional to how well you did, where every step up is worth the same as the last.
Now notice what it does to a gamble. Take a roster that is reliably decent, finishing somewhere between fifth and seventh nearly every time. Its expected payout is zero, because none of those finishes pay. Now take a wilder team that finishes first one year in six and in the bottom half the rest. Its average finish is worse; its expected payout is far better, because the one outcome that pays is one it can actually reach. The reliable team is maximising its average finish in a contest that does not pay for average finishes.
Why the downside is already floored
The reason variance is free here, rather than merely tempting, is that the loss is capped before you start. You paid your stake. If your season goes badly, you lose the stake. If it goes catastrophically — every pick busts, you finish dead last — you lose exactly the same stake. Losing by a lot is not worse than losing by a little. There is no additional penalty waiting below.
So anything that widens the spread of possible seasons adds value at the top and costs nothing at the bottom, because the bottom is a floor rather than a slope. This is the same reason a call option is worth more than the expected value of the thing it tracks: the holder keeps the whole upside and hands the entire downside to the floor.
Which is why "he might bust" is, on its own, not an objection to a pick. Busting is already priced in and already survivable. The real objection is that a pick might be unremarkable — a player who reliably delivers decent, replaceable output is offering exactly the outcome the payout does not reward.
The qualification that makes this safe
If the note stopped there it would be dangerous advice, because pure variance never reaches the money at all. A real fantasy season is a two-stage tournament, and the two stages reward opposite things.
Stage one is a fourteen-week qualifying campaign in which you must finish in the top half of the league to advance at all. What survives fourteen accumulated head-to-head results is a high mean — a team that scores well most weeks qualifies. A boom-or-bust team that posts four spectacular weeks and ten poor ones does not, no matter how large the four were. The bracket cannot be reached by variance.
Stage two is a three-game bracket in which accumulation stops entirely. The task is to beat one specific opponent in one specific week, three times. Average scoring becomes irrelevant the moment you qualify; what matters is the size of the good weeks. A team capable of a huge day beats a steady team roughly half the time, and half of three coin flips is far more than a steady team's realistic share.
So the correct sequencing is: the regular season buys the ticket, the playoffs cash it. You cannot buy the ticket with variance and you cannot cash it with consistency. See How the league is won for this argument in its concrete, this-league form.
The resulting shape
The resolution is not a compromise. It is a rule about where in the roster each side applies.
Protect the mean where the picks are expensive and the field is thin. The earliest picks buy the qualifying campaign: they start every week for fourteen weeks, they are the largest single input into whether the average clears the cut, and they come from a shallow supply that the waiver wire cannot replace. This is the wrong place to be clever. How much to spend on running backs is the position-specific version.
Buy ceiling where it is cheap and safety is worth least. Later, the reliable players are gone and what remains is priced on expectation, which makes upside inexpensive. More importantly, a safe bench player is the least valuable thing you can own. A reserve you can trust for a modest, predictable output supplies exactly what the free-agent pool gives away, and it will never win a playoff game. The bench exists to hold lottery tickets, not spare parts.
What "buying variance" means at the roster level
Concretely, in the part of the draft where the rule applies:
- Prefer high-ceiling, low-floor players over steady mediocrity. A player whose weekly scores swing widely contributes exactly what the bracket rewards, and the weeks he disappoints cost a bench slot rather than a season.
- Prefer upside to insurance. A backup rostered only to protect a starter you already own converts a bench slot into a hedge, and a hedge lowers variance by construction. Some hedging is prudent; systematic hedging is buying the wrong product.
- Avoid pure floor plays. Filling the last roster spots with the most predictable available bodies optimises for the outcome that pays nothing — which is also why those slots should not go to positions where the best option is barely distinguishable from the twelfth (The last two rounds — kicker and defence).
- Do not confuse variance with carelessness. Buying variance means choosing wide distributions on purpose, at a discount. It does not mean ignoring role or taking players whose ceiling is low and whose floor is low. A wide distribution is a feature; a low one is not.
Open questions
- The crossover round is asserted, not measured. That the mean matters most early and ceiling most late is sound in direction, but the pick at which the balance tips is a judgement call here, not a computed answer. Measuring it would need a simulation of full seasons under this league's exact settings and payout, and nobody has run one.
- The advantage a high-variance team holds in a three-game bracket is real, but its size depends on the spread of weekly scores among the qualifying teams, which is not observable until a season is well underway (Sample size in a 17-game season).
- Treating the qualifying stage as purely mean-driven is a simplification: the schedule is randomly generated, so qualifying already contains substantial luck, and a fully correct treatment would trade mean against variance in both stages.