A live sportsbook risk desk racing three online sign-assignment strategies against a stream of random exposure vectors.
Grounded in arXiv:2608.10040 โ "Online Discrepancy Minimization for Sub-Gaussian Inputs via Regularization and Restriction."
The paper studies the online vector balancing (discrepancy) problem: sub-Gaussian vectors arrive one at a time and must be
irrevocably signed ยฑ1 so the running signed sum stays as close to 0 as possible. The paper's algorithm
regularizes the Lโ norm with a smooth, two-sided exponential ("hedge") potential in place of the hard max, and
restricts each sign decision to an adaptively chosen coordinate set โ the markets currently carrying the most risk โ
rather than weighing every coordinate equally. That combination provably keeps the running discrepancy far below the
O(โn) blow-up that a naive greedy or random sign choice is stuck with. Below, each incoming vector is a book's
simultaneous exposure across 6 betting markets; each lane is a different signing strategy; the lane's readout is its
running liability = โSโโ, the single most-exposed market.
6/s
0.35
3 / 6
round 0
seed โ
Incoming exposure vector vt (|bet size| per market, $k)
Running liability โStโโ โ lower hugging zero is better
Regularize + Restrictthe paper's algorithm
$0.0k
Minimizes a two-sided exponential ("hedge") potential ฮฃ cosh(ฮทยทS_i) over only the coordinates currently carrying the most risk โ an adaptively restricted coordinate set โ instead of the full vector.
Pure Greedybaseline
$0.0k
Picks whichever sign locally minimizes the instantaneous โS_tโโ this round, with no memory of trend or regularization. Myopic โ can still drift.
Random Bookmakerno strategy
$0.0k
Flips a fair coin for every vector, completely ignoring exposure. The textbook random-walk discrepancy, which grows like โn.