{"id":"75585358-ff3e-4de6-ac77-3e9d16788397","arxiv_id":"2505.21543","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Boltzmann-transformed betting odds, combined with team history via a Dirichlet prior, are claimed to outperform raw odds in a Kelly betting simulation, but the evaluation is statistically and methodologically flawed.","lead":"A short paper proposes turning soccer betting odds into imaginary energy levels and then applying the Boltzmann formula to get new probabilities, claiming these beat the raw odds in a five-season betting simulation. The method is a hand-made nonlinear transform with no statistical justification, and the evaluation uses a mis-specified Kelly criterion, so the central claim is not supported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed edge is an artifact of a mis-specified Kelly criterion and profit calculation: winning bets are credited with stake plus profit instead of profit alone, and the 'Kelly' formula is not the Kelly fraction, so the comparison is not a valid test of probability estimation.","rationale":"The reader's verdict (REJECT) is correct, but the single most load-bearing concern is not the Dirichlet pseudo-count assumption; it is the evaluation metric itself. The paper's only evidence for 'consistently outperform' is the Kelly simulation, and that simulation is mathematically wrong in two ways that systematically favor the Boltzmann probabilities. The Dirichlet update is indeed invalid as a Bayesian posterior because the count vector C mixes wins by two different teams and an averaged draw count, but the algorithm could still be treated as an ad hoc heuristic; if the simulation were correct, the empirical comparison might still be informative. However, the incorrect Kelly formula and the stake-as-profit error strike at the very definition of the outcome measure. Correcting them could reverse the edge, which would leave the central claim with no supporting evidence. The concrete test is straightforward and decisive: recompute the gains with the correct formulas. We therefore keep the reader's REJECT verdict (UNCHANGED), while noting that the weakest link is the evaluation protocol rather than the Bayesian semantics.","tokens_in":7616,"tokens_out":6949,"duration_ms":63855,"concrete_test":"Reimplement Pseudocode 2 and 3 with the correct fractional Kelly fraction f * (p*O - 1)/(O - 1) and the correct profit formula stake*(O - 1) on wins, using normalized baseline probabilities (p = (1/O)/sum(1/O)) for the original odds. Recompute GainBoltz - GainOriginal over the five seasons. If the cumulative edge drops below zero or fails a bootstrap significance test across matches/seasons, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is supported only by the betting simulation in Sections 4.2 and 4.3 (Pseudocode 2 and 3), but that simulation uses an incorrect Kelly criterion and an incorrect gain calculation. Step 2 defines KELLY_O = f * (PostO - (1 - PostO)/OddsO). The correct fractional Kelly stake for a bet with decimal odds O and win probability p is f * (p*O - 1)/(O - 1). The paper's expression is larger by approximately f*p/(O - 1) and can be positive even when the bet has negative expected value (e.g., p = 0.75, O = 1.33 gives KELLY > 0 under their formula, while p*O - 1 < 0). Step 3 then computes GAIN_O = 100 * KELLY_O * OddsO on a win. With decimal odds, a stake S yields profit S*(OddsO - 1), not S*OddsO; the paper credits the stake itself as profit, overstating every winning bet's gain by the stake amount. Because the Boltzmann transformation pushes favorite probabilities toward 1 (Section 3.3), it increases KELLY_O and therefore the stake, so both errors disproportionately inflate the Boltzmann totals. The reported $1,841.57 edge may be entirely an accounting artifact rather than evidence of improved probability estimates. This invalidates the central claim regardless of whether the Dirichlet update (the reader's concern) is statistically meaningful.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Boltzmann-informed probabilities' for soccer match outcomes by defining energy levels from betting odds, converting them via the Boltzmann distribution, using the result as a Dirichlet prior over historical win/draw counts, and then evaluating the posterior through a Kelly-criterion betting simulation on five English Premier League seasons. The central claim is that this procedure consistently outperforms probabilities derived directly from the odds, with a cumulative gain edge of $1,841.57.","tokens_in":7984,"tokens_out":5806,"duration_ms":48582,"significance":"If correct, the approach would offer a simple, physics-inspired correction to market-implied probabilities and would be of interest to complex-systems modeling. The manuscript is transparent: it provides pseudocode for every step, uses a concrete five-season dataset, and reports season-by-season results, which makes the claims reproducible in principle. However, the evaluation rests on a mis-specified Kelly criterion, an incorrect profit calculation, and an invalid Dirichlet posterior, so the headline result is not a valid test of probability estimation.","major_comments":[{"comment":"The stake formula KELLY_O = f·(PostO − (1 − PostO)/OddsO) is not the Kelly criterion for decimal odds; the correct fraction is f·(PostO·OddsO − 1)/(OddsO − 1). As written, the formula yields a positive stake whenever PostO > 1/(OddsO + 1), which includes many negative expected-value bets (e.g., OddsO = 2, PostO = 0.4). Consequently, the simulation places bets without a positive edge, and the reported gains cannot be attributed to superior probability estimates.","section":"§4.2, Pseudocode 2, Step 2"},{"comment":"A winning bet is credited with GAIN_O = 100·KELLY_O·OddsO, but with decimal odds the stake is returned along with the winnings, so the profit is 100·KELLY_O·(OddsO − 1). This overstates every winning bet by the stake amount, and since the Boltzmann transform inflates favorite probabilities and hence KELLY_O for favorites, the inflation is systematically larger for the Boltzmann-informed model. The $1,841.57 edge is therefore at least partly an accounting artifact.","section":"§4.3, Pseudocode 2, Step 3"},{"comment":"The vector C = [WH, D, WA] is treated as a multinomial count vector for the current match in a Dirichlet prior/posterior update, but WH and WA are the historical win totals of the two teams and D is the average of their draw counts; these are not mutually exclusive outcomes of a single exchangeable trial. The resulting posterior is not a valid Bayesian posterior for the match outcome distribution, so the probabilities PostH, PostD, and PostA that drive the simulation are not statistically meaningful.","section":"§3.4–§3.5 and Pseudocode 1"}],"minor_comments":[{"comment":"The max-entropy example uses arbitrarily chosen energy values and then asserts that the Boltzmann distribution is the least biased distribution for that average energy; this is true for any energy assignment and does not justify the particular odds-derived energy mapping used in the paper.","section":"§3.3"},{"comment":"The original odds probabilities PHome, PDrew, and PAway are taken as 1/Odds without normalization, whereas the baseline percentages in §3.1 (32%, 22%, 46%) are normalized; this inconsistency could affect the 'original probabilities' baseline in the Kelly simulation.","section":"Pseudocode 3"},{"comment":"There are several typographical issues, including 'PDrew' for 'PDraw', 'Drew' for 'Draw', 'bank role' for 'bankroll', and 'accommodating edge' for 'accumulated edge' in the Results section.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is invalidated by the mis-specified Kelly criterion and the incorrect gain calculation, which are not mere presentation issues but affect every reported number in Tables 1 and 2. Even if those were corrected, the Dirichlet posterior is not a valid Bayesian update, so the probability estimates lack statistical grounding. The manuscript is transparent and the topic is interesting, but the core result is an artifact of these errors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one genuinely new thing here is using a Boltzmann (softmax) transform on hand-defined energy functions of betting odds, then treating the output as a Dirichlet prior for match outcomes. That is a concrete, testable heuristic and it is honestly presented as an illustration, not as a physics claim. The paper is clearly written and the pseudo-code helps. It does not, however, establish what it claims.\n\nThe load-bearing problem is the evaluation. The Kelly fraction is not f*(p - (1-p)/O). That is not the Kelly criterion for decimal odds; it can be positive when the bet has negative expected value. On top of that, a winning bet is credited with stake plus profit instead of profit alone. With decimal odds, a stake S returns profit S*(O-1), not S*O. Both errors inflate returns disproportionately for the Boltzmann probabilities, which push favorite probabilities toward 1 and therefore increase the stake. The reported $1,841.57 edge is very plausibly an accounting artifact rather than evidence of better probability estimates. The stress-test note is right, and it lands exactly on the central comparison in Sections 4.2 and 4.3.\n\nThe reader's other main concern also holds: the Dirichlet update treats [WH, (DH+DA)/2, WA] as multinomial counts for a single match. Those counts do not come from one match, they come from two teams' histories, so the \"posterior\" is not a legitimate Bayesian posterior for the match outcome. That undermines the statistical interpretation even if the Kelly errors were fixed.\n\nWhat should be credited: the paper does not hide its energy definitions as derived from first principles; it calls them heuristic. It also includes a control (GainOriginal+Bayes) to separate the Boltzmann effect from the Bayesian update. That control is a good instinct, though it inherits the same accounting errors. The max-entropy discussion is fine as motivation but does not justify the specific energy choices, which remain arbitrary. No data or code are provided, so the exact numbers cannot be checked.\n\nNet: the central claim is not credible as written. The idea could be revisited with a correct Kelly formula, a proper predictive model, and significance testing, but this paper does not get there. A serious editor could send it to peer review because the claim is concrete and the flaws are instructive, but I would expect rejection. My own recommendation is to treat it as a desk reject if the journal is selective; the empirical result is an artifact and the method section does not salvage it.","headline":"The empirical claim is an artifact of a mis-specified Kelly criterion and an incorrect profit calculation; the underlying idea is not crazy but the paper as written does not support it.","tokens_in":8444,"tokens_out":918,"would_cite":false,"duration_ms":7056,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Boltzmann-informed probabilities, built from odds-derived energy levels, beat raw betting odds in Kelly betting over five English Premier League seasons, with a $1,841.57 cumulative edge.","keywords":["Boltzmann distribution","Kelly criterion","sports betting","English Premier League","energy-informed probabilities","Dirichlet prior","maximum entropy","favorite-longshot bias"],"falsifier":"Recompute the five-season Kelly simulation using a valid multinomial likelihood—for each match, use counts of the actual outcomes of previous matches involving the two teams as the Dirichlet observation vector, instead of $C=[W_H,D,W_A]$. If the cumulative edge over the original odds falls materially below $1,841.57 or reverses sign, the reported improvement is an artifact of the invalid count vector rather than a property of Boltzmann-informed probabilities.","tokens_in":7411,"feed_emoji":"⚽","tokens_out":9822,"duration_ms":75066,"temperature":0.7,"pith_summary":"The paper proposes that assigning hypothetical energy levels to the possible outcomes of a random variable, then feeding those energies through Boltzmann's exponential distribution, can produce better probability estimates than the raw probabilities. It tests this idea on five English Premier League seasons, translating each match's betting odds into energy levels, using the resulting Boltzmann distribution as a prior, and combining it with historical team counts in a Dirichlet posterior update. When those posterior probabilities are used to size bets through the Kelly criterion, they consistently beat bets sized with the original odds, accumulating a $1,841.57 edge over the five seasons. The paper argues that the Boltzmann transform corrects behavioral biases such as the favorite–longshot bias by exponentially sharpening the gap between favorites and underdogs, and that energy-based probability modeling may transfer to other complex social systems.","feed_headline":"Boltzmann odds beat raw odds by $1,841 in Kelly betting","feed_subtitle":"Five seasons of Premier League data give Boltzmann-informed probabilities a $1,841.57 edge in Kelly-criterion betting.","key_machinery":"The load-bearing device is the mapping from betting odds to energy levels, $E_H=O_H/O_A$, $E_D=O_D$, $E_A=O_A/O_H$, followed by the Boltzmann distribution $p_i=e^{-E_i}/Z$. This transform is what sharpens the odds: it converts a favorite's small implied probability into a tiny Boltzmann weight and an underdog's large implied probability into an even smaller one, so the favorite's posterior probability is amplified. The resulting probabilities are not used as final predictions but as the $\\alpha$ parameters of a Dirichlet prior; the posterior is obtained by adding the historical count vector $C=[W_H,D,W_A]$. The evaluation machinery is the fractional Kelly criterion, which decides bet size from the gap between the posterior probability and the odds.","core_discovery":"The central claim is that a probability estimate built from a Boltzmann distribution over odds-derived energy levels is more accurate than the probabilities the odds imply directly. For a match with average odds $O_H,O_D,O_A$, the paper defines $E_H=O_H/O_A$, $E_D=O_D$, $E_A=O_A/O_H$, then sets $p_i=e^{-E_i}/Z$ with $Z=\\sum_j e^{-E_j}$. These Boltzmann probabilities are justified as the maximum-entropy distribution under the constraint of the average energy. They are not used as final predictions; they set the $\\alpha$ parameters of a Dirichlet prior that is combined with a historical count vector $C=[W_H,D,W_A]$ (where $D$ is the average of the two teams' draw counts) to produce posterior probabilities for the match. Feeding those posterior probabilities into a fractional Kelly criterion with $f=0.2$ and $100 per match yields a cumulative gain of $9,214.74 over five English Premier League seasons, versus $7,373.17 for the same betting rule using the original odds, an edge of $1,841.57. A control that applies the same Bayesian update to the original probabilities earns only $5,222.73, which the paper reads as evidence that the Boltzmann transform, not the Bayesian machinery, produces the improvement.","pith_inferences":["The count vector $C=[W_H,D,W_A]$ does not describe outcomes of a single multinomial trial, so the 'posterior' is not a valid Bayesian posterior for the match; an evaluation that replaces it with counts of actual previous match outcomes would show whether the edge survives a correct likelihood.","The Boltzmann transform is a monotone sharpening of the odds, so the reported edge may be reproducible by any sufficiently strong odds transform; comparing against a fitted softmax with a free temperature would test whether Boltzmann's specific exponential form matters.","The simulation's bet sizing parameters—$f=0.2$ and a $100 per-match bankroll—are arbitrary; sensitivity analyses over these parameters and over bookmaker margins would indicate whether the cumulative edge is robust."],"forward_implications":["If the edge is real, any market that publishes odds or prices for mutually exclusive outcomes can be recalibrated with the same parameter-free energy transform, including political and financial prediction markets.","The control result—original odds plus the same Bayesian update earning less than both alternatives—implies that the improvement is attributed to the Boltzmann transform itself, not to the Dirichlet smoothing.","Because the transform amplifies the distance between favorites and underdogs, applying it should counteract the documented favorite–longshot bias in betting markets.","The same energy-based modeling recipe is already pointed toward armed-conflict fatality prediction, suggesting the paper intends the method as a general tool for complex social systems, not just soccer."],"supporting_citations":[{"why":"Introduces the energy-informed probabilistic approach that the paper extends from extreme events to match outcomes.","marker":"[1]"},{"why":"Companion under-review model that applies the same Boltzmann-informed framework to conflict fatalities and is cited as a source of the method.","marker":"[2]"},{"why":"Establishes betting odds as a competitive forecasting instrument, the baseline the Boltzmann-informed probabilities are measured against.","marker":"[3]"},{"why":"Defines betting market efficiency and market prices as best forecasts, the inefficiency the paper claims to correct.","marker":"[4]"},{"why":"Supplies the behavioral-bias account (overbetting popular teams, herd distortions) that the method is said to filter.","marker":"[5]"},{"why":"Documents the favorite–longshot bias that the Boltzmann transform is invoked to counteract.","marker":"[6]"},{"why":"Provides the maximum-entropy principle used to justify the Boltzmann distribution as the least-biased distribution under an energy constraint.","marker":"[7]"}],"fun_headline_variants":["Boltzmann probabilities edge out raw odds by $1,841 in Kelly betting","Energy-based probabilities beat odds in Premier League betting test","Boltzmann odds gain $1,841 over raw in Kelly criterion study","Boltzmann-informed betting wins $1,841 more than raw odds","Physical energy concept improves sports betting probability estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the historical counts $C=[W_H,D,W_A]$, where $D$ is the average of the two teams' draw counts, can be treated as a multinomial count vector in a Dirichlet posterior update for the current match; if that premise fails, the posterior probabilities that drive the Kelly simulation are not statistically meaningful and the comparison to raw odds is invalid.","fun_headline_variants_meta":{"raw":{"variants":["Boltzmann probabilities edge out raw odds by $1,841 in Kelly betting","Energy-based probabilities beat odds in Premier League betting test","Boltzmann odds gain $1,841 over raw in Kelly criterion study","Boltzmann-informed betting wins $1,841 more than raw odds","Physical energy concept improves sports betting probability estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1289,"prompt_tokens":898,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":305}},"tokens_in":514,"tokens_out":391,"duration_ms":3126,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:23:47.296752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the five-season Kelly simulation using a valid multinomial likelihood—for each match, use counts of the actual outcomes of previous matches involving the two teams as the Dirichlet observation vector, instead of $C=[W_H,D,W_A]$. If the cumulative edge over the original odds falls materially below $1,841.57 or reverses sign, the reported improvement is an artifact of the invalid count vector rather than a property of Boltzmann-informed probabilities.","supporting_citations":[{"cited_title":"Extreme events in armed conflicts: a perspective","cited_arxiv_id":null,"evidence_quote":"Introduces the energy-informed probabilistic approach that the paper extends from extreme events to match outcomes."},{"cited_title":"and Cohen, Y ., under review","cited_arxiv_id":null,"evidence_quote":"Companion under-review model that applies the same Boltzmann-informed framework to conflict fatalities and is cited as a source of the method."},{"cited_title":"and Skiera, B., 2009","cited_arxiv_id":null,"evidence_quote":"Establishes betting odds as a competitive forecasting instrument, the baseline the Boltzmann-informed probabilities are measured against."},{"cited_title":"and De Angelis, L., 2019","cited_arxiv_id":null,"evidence_quote":"Defines betting market efficiency and market prices as best forecasts, the inefficiency the paper claims to correct."},{"cited_title":"Betting Against the Crowd: A Complex Systems Approach","cited_arxiv_id":null,"evidence_quote":"Supplies the behavioral-bias account (overbetting popular teams, herd distortions) that the method is said to filter."},{"cited_title":"and Whelan, K., 2025","cited_arxiv_id":null,"evidence_quote":"Documents the favorite–longshot bias that the Boltzmann transform is invoked to counteract."}],"review_version":1}