Betting Loss: Theory and Applications

• Betting loss is a cumulative metric quantifying exposure via repeated betting decisions, linking negative returns, ruin events, and bookmaker liabilities.
• The framework employs multiplicative wealth processes, martingale analysis, and convex duality to derive variance-adaptive bounds and robust statistical estimators.
• Empirical and theoretical studies show that misallocated stakes, incoherent odds, and strategic risk management can lead to sure losses or sublinear excess liability.

Searching arXiv for recent and relevant papers on betting loss and related betting-based methods. Betting loss is not a single object but a family of mathematically related notions that arise in gambling theory, sequential concentration, learning theory, and market design. Across recent arXiv literature, it can denote a bettor’s negative return or ruin event, a bookmaker’s worst-case payout burden, a bookmaker’s “sure loss” under incoherent odds, or a betting-inspired surrogate objective for statistical estimation (Nakharutai et al., 2019, Tal et al., 19 Jun 2025, Li et al., 16 Jul 2025). A common structural motif is multiplicative wealth or liability evolution under repeated decisions: gains and losses are accumulated pathwise, and analysis proceeds through martingales, convex duality, chance constraints, or minimax dynamic programming. In a further theoretical extension, Kolmogorov-Loveland betting strategies are shown to “lose the Betting game on open sets,” meaning that no finite family of such strategies can force arbitrarily large capital on every sequence in every sufficiently small effective open set (Petrović, 2024).

1. Term, scope, and core quantities

In bettor-centric work, betting loss is usually measured directly in monetary or wealth terms. Examples include return on investment,

$$ROI = \frac{\text{Total Return} - \text{Total Stake}}{\text{Total Stake}},$$

capital loss as $-ROI$, the maximal bet size $M$, the maximal deficit, and ruin probability

$f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$

depending on the model under study (Makinde et al., 9 Apr 2026, Zubrilina, 2018, Roy et al., 11 Dec 2025). In bookmaker-centric work, loss becomes liability: under fixed odds it may be the existence of a nonnegative combination of gambles that forces the bookmaker into sure loss, whereas in online bookmaking it is the minimax total payout

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$

against adversarial bettors and outcomes (Nakharutai et al., 2019, Tal et al., 19 Jun 2025).

A distinct usage appears in statistical learning. For $[0,1]$-valued regression, “betting loss” is a novel loss function constructed from log-wealth increments, and its minimizer yields a variance-adaptive second-order generalization bound without explicit variance modeling (Li et al., 16 Jul 2025). In sequential concentration, betting enters not as a loss function but as a proof device: a nonnegative wealth process with a pathwise lower bound in terms of

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$

is converted into a time-uniform martingale deviation inequality and then into an empirical Bernstein law of the iterated logarithm (Orabona, 21 May 2026).

These usages are conceptually linked by the same asymmetry. Small, frequent gains may coexist with rare but catastrophic liabilities; conversely, carefully controlled betting fractions can generate concentration guarantees or robust estimation procedures. This suggests that “betting loss” is best understood as a structural notion of cumulative exposure rather than as a single application-specific metric.

2. Betting loss as a regression objective

For i.i.d. $[0,1]$-valued regression with realizability,

$\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$

the standard target is small expected absolute prediction error

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$

The classical squared-loss minimizer gives a worst-case-style guarantee, while the log-loss minimizer yields a first-order bound scaling with $-ROI$0 rather than a purely worst-case variance proxy (Li et al., 16 Jul 2025). The reported log-loss guarantee is

$-ROI$1

The betting loss is introduced to obtain a strict second-order improvement. For clipping parameter $-ROI$2 and scale parameter $-ROI$3,

$-ROI$4

and the empirical objective is

$-ROI$5

The estimator is

$-ROI$6

The log term is explicitly interpreted as a log-wealth increment of a gambler betting on the residual $-ROI$7; the maximization over $-ROI$8 induces a robust minimax geometry (Li et al., 16 Jul 2025).

The main guarantee is variance-adaptive. Writing

$-ROI$9

the paper states that with probability at least $M$0, for all $M$1,

$M$2

For $M$3, the excess term vanishes, giving

$M$4

The result is termed variance-adaptive because no knowledge of the variance is used by the algorithm, in contrast to explicit modeling of label variance or the label distribution itself as part of the function class (Li et al., 16 Jul 2025).

The same betting perspective also sharpens PAC-Bayes concentration. A coin-betting PAC-Bayes inequality controls posterior averages of the exact optimal log-wealth

$M$5

simultaneously for all sample sizes, and earlier Bernoulli-KL and empirical Bernstein PAC-Bayes bounds are recovered as relaxations (Jang et al., 2023). This suggests a broader role for betting loss constructions: they preserve exact log-wealth structure longer than quadratic surrogates do, and the retained structure translates into tighter statistical guarantees.

3. Wealth processes, concentration, and LIL behavior

A central betting-based construction begins with outcomes $M$6. At round $M$7, the bettor chooses $M$8 based on $M$9 and receives payoff $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$0, so

$f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$1

A key representation is a mixture over constant fractional bets $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$2: if $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$3 is a distribution over $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$4, then

$f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$5

and hence

$f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$6

Because $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$7 whenever $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$8 and $f(x,p):=P(W_n\le 0\text{ for some }n\mid W_0=x),$9, the wealth process is nonnegative (Orabona, 21 May 2026).

The specific prior used in the empirical Bernstein LIL analysis is symmetric around $$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$0 with density proportional to

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$1

for $$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$2. The empirical optimizer is

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$3

and the key convex-analytic quantity is

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$4

A related conjugacy calculation yields

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$5

which is what allows interpolation between Gaussian-like $$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$6 behavior and linear-in-$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$7 logarithmic corrections (Orabona, 21 May 2026).

The resulting lower bound on wealth has leading term

$$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$8

When $$\mathcal{L}_{T,K}=\inf_{\Psi^T}\max_{q^T\in\Delta^T}\max_{k\in[K]}\sum_{t=1}^T \frac{q_t(k)}{r_t(k)}$$9 is a martingale difference sequence with $[0,1]$0, the wealth process is a martingale under any predictable betting strategy, and Doob’s maximal inequality gives

$[0,1]$1

Combining maximal control with the pathwise lower bound yields a uniform-in-time inequality, which can then be inverted through

$[0,1]$2

and

$[0,1]$3

A readable deviation form is

$[0,1]$4

where

$[0,1]$5

This is an empirical Bernstein-style law of the iterated logarithm: finite-time, data-dependent, self-normalized, and uniform over time (Orabona, 21 May 2026).

The same broad template appears in PAC-Bayes coin-betting. There, the exact object is the posterior average of optimal log-wealth rather than a relaxed quadratic expression, and confidence sequences are obtained by retaining this exact wealth formulation until the final inversion step (Jang et al., 2023). A plausible implication is that “betting loss” methods are not merely analogies to gambling; they are a reusable analytical mechanism for deriving sharp time-uniform inequalities from nonnegative wealth processes.

4. Ruin, bankroll dynamics, and staking systems

In classical gambling models, betting loss is often dominated by path-dependent staking rules rather than by unfavorable expected value alone. The Labouchere system is exemplary. Starting from a list $[0,1]$6 of positive numbers, the player bets

$[0,1]$7

deleting the first and last entries on a win and appending the lost bet on a loss. The maximal bet is

$[0,1]$8

For $[0,1]$9, where $$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$0 solves

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$1

the expectation $$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$2 is finite, partially answering a question of Grimmett and Stirzaker (Zubrilina, 2018). A related simulation study describes the same system as psychologically appealing because it may produce apparently consistent linear returns, yet it contains a “fatal flaw”: after a losing streak the next required wager can exceed available capital, and with continued play the bankroll balance is described as tending toward zero (Billings et al., 2017).

A different reactive rule reverses familiar martingale intuition. If

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$3

then the next bet doubles after a win and halves after a loss. The ruin probability

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$4

satisfies the exact threshold

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$5

Thus survival with positive probability occurs only when the game is unfavorable and the initial fortune exceeds $$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$6 (Roy et al., 11 Dec 2025). The paper derives the functional equation

$$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$7

shows that $$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$8 is increasing and real-analytic in $$M_t:=\sum_{i=1}^t g_i,\qquad S_t:=\sum_{i=1}^t g_i^2$$9, and proves that $[0,1]$0 is singular continuous and Hölder continuous (Roy et al., 11 Dec 2025). The underlying mechanism is that wins increase stake size and thereby magnify later exposure.

Kelly-style strategies illustrate a complementary point: positive expected growth does not preclude severe realized loss. In feedback-control form,

$[0,1]$1

and survivability is exact: $[0,1]$2 A Taylor approximation yields

$[0,1]$3

but this may lie outside the survivable interval; in the paper’s example, $[0,1]$4 while the admissible interval is $[0,1]$5, so a single worst-case loss drives wealth negative (Hsieh, 2020). Uncertainty in probability estimates creates a related overbetting problem. In horse-race Kelly betting with logistic-normal uncertainty, the standard plug-in model reduces total expected return from $[0,1]$6 under true probabilities to $[0,1]$7, while a combined expectation-plus-chance-constraint model reaches $[0,1]$8 (Metel, 2017).

Empirical work on human play shows that even favorable odds do not prevent betting losses when stake sizing is poor. In a 60/40 biased-coin experiment with $[0,1]$9 quantitatively trained participants, $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$0 subjects went all in on a single flip, $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$1 subjects bet on tails at some point despite the stated bias toward heads, about $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$2 went bust, more precisely $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$3 received no payout, only $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$4 reached the $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$5 (Haghani et al., 2017). In exotic horse wagering, the response is to impose a chance constraint on the probability of obtaining at least threshold return $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$6 within horizon $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$7: $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$8 The corresponding mixed-integer nonlinear program trades off Kelly-style growth against losing-streak control, and shorter tolerated streaks empirically reduce return while making the strategy more selective (Deza et al., 2015).

5. Market behavior, promotions, and observed bettor losses

A bookmaker’s offers can be represented as a set of desirable gambles, and the coherence question becomes whether this set avoids sure loss. For odds $\mathbb E[y\mid x]=f^*(x),\qquad f^*\in\mathcal F,$9 on outcome $$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$0, the bookmaker’s gamble is

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$1

A collection $$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$2 avoids sure loss if

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$3

If this fails, the bookmaker incurs sure loss and the customer has sure gain (Nakharutai et al., 2019). With fixed fractional odds, the induced upper probability is

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$4

and the odds avoid sure loss iff

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$5

Free coupons can break this coherence. The paper models “first bet + free second bet” through a derived gamble $$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$6, computes $$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$7 by the Choquet integral, and interprets any negative value as exploitable. In the simplified Forest example,

$$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$8

so the customer can secure a gain of $$\mathbb E_{x\sim\mathcal D_X}\bigl[|f^*(x)-\hat f(x)|\bigr].$$9 pounds (Nakharutai et al., 2019).

Behavioral betting losses need not arise from arbitrage failure; they may instead reflect systematically biased stake allocation. In Bundesliga in-play betting after an equalizing goal at $-ROI$00, the equalizer coefficient in the bettor stake regression is

$-ROI$01

meaning $-ROI$02 percentage points higher relative stake on the team that scored the equalizer. The paper also reports $-ROI$03 higher stakes at controls’ means, $-ROI$04 in another specification, and $-ROI$05 percentage points in the second half (Ötting et al., 2022). Yet the equalizer coefficient is insignificant in both the outcome regression and the odds regression, and the average overround after the equalizer is $-ROI$06. The reported ROI from always backing the equalizing team is $-ROI$07 for strong favourites, $-ROI$08 for moderate favourites, $-ROI$09 for moderate longshots, and $-ROI$10 for strong longshots, leading the paper to conclude that betting on apparent momentum would lead to substantial losses (Ötting et al., 2022).

Not all empirical work finds persistence after losses. In mobile sports betting among over $-ROI$11 Kenyan smartphone users, gamblers are less likely to bet following poor results and more likely to bet following good results, with nearly symmetric magnitudes: a positive feedback shock raises the probability of betting by about $-ROI$12 percentage points, while a negative feedback shock lowers it by about $-ROI$13 percentage points (Blumenstock et al., 2020). The paper interprets this as consistent with Bayesian updating and finds no evidence that increased betting leads to increased debt under its instrumental-variables design (Blumenstock et al., 2020). This suggests that loss-chasing is not the only possible behavioral response; learning from losses is empirically observable in some settings.

Social-media-mediated betting adds another layer. Tracking $-ROI$14 verified pre-match bets from three Nigerian tipsters over approximately $-ROI$15 million in tracked volume, one study reports that the influencers themselves collectively lost $-ROI$16 on their promoted bets, while a flat-staking follower would lose $-ROI$17 (Makinde et al., 9 Apr 2026). The overall win rate is $-ROI$18, low odds $-ROI$19 still lose about $-ROI$20, and high odds $-ROI$21 lose about $-ROI$22. Fixed Return performs best among the tested staking strategies, followed by Inverse, then Square Root, with Flat worst, but none are profitable (Makinde et al., 9 Apr 2026). At a more structural level, the two-armed Futurity bandit formalizes the slogan “long bet will lose”: even if each arm is individually calibrated to look fair, the combined two-armed system yields asymptotic casino profit

$-ROI$23

with equality only when $-ROI$24 (Chen et al., 2022). A plausible implication is that many observed betting losses are generated not by isolated bad forecasts but by architectures that are fair-looking locally and unfavorable globally.

6. Bookmaker loss and adversarial online odds

In online bookmaking, betting loss is the bookmaker’s worst-case total payout burden in a repeated game where odds can be updated sequentially. With $-ROI$25 outcomes and $-ROI$26 rounds, the bookmaker chooses $-ROI$27, the gambler chooses $-ROI$28, and if outcome $-ROI$29 realizes then round-$-ROI$30 liability is

$-ROI$31

The total worst-case loss is therefore

$-ROI$32

This quantity determines net gain under overround $-ROI$33: after collecting $-ROI$34 betting units, the bookmaker keeps

$-ROI$35

The exact solution is that $-ROI$36 is the largest real root of

$-ROI$37

(Tal et al., 19 Jun 2025).

For $-ROI$38, the polynomial reduces to

$-ROI$39

whose largest root is

$-ROI$40

This is precisely the binary solution obtained via bi-balancing trees (Bhatt et al., 12 Jan 2025). In that formulation, decisive gamblers $-ROI$41 are the worst-case adversaries, and the optimal strategy equalizes house loss across all decisive paths. The guaranteed bookmaker gain is

$-ROI$42

so positive guaranteed profit is possible iff

$-ROI$43

(Bhatt et al., 12 Jan 2025).

The many-outcome theory sharpens this picture. The Bellman value function is

$-ROI$44

and the paper explicitly characterizes the Bellman-Pareto frontier. The corresponding optimal opportunistic bookmaking loss from state $-ROI$45 is again a largest root of a state-dependent polynomial, and the optimal odds are given by a closed form involving denominator polynomials $-ROI$46 (Tal et al., 19 Jun 2025). The worst-case loss is attained against decisive gamblers, but when the gambler is non-decisive the bookmaker can do better than the minimax benchmark by recomputing the opportunistic continuation value (Tal et al., 19 Jun 2025).

Asymptotically, for fixed $-ROI$47,

$-ROI$48

so

$-ROI$49

The leading regret coefficient

$-ROI$50

is the largest root of the $-ROI$51-th probabilist’s Hermite polynomial $-ROI$52 (Tal et al., 19 Jun 2025). In the paper’s terminology, this means bookmakers can be “as fair as desired” while avoiding financial risk. More precisely, fairness here means that the excess liability above collected stakes is sublinear, not that the market is neutral in any behavioral or welfare sense.