How to Predict the Future (Mathematically)
Credibility score: 88/100 — Highly Credible. This video is highly credible with well-supported claims.
BSmeter analyzed "How to Predict the Future (Mathematically)" and rated it 88/100 for credibility (a BS score of 12/100 — highly credible), on 2026-04-07. Its weakest claim — "Coin flip prediction game: $1 per correct guess over M rounds, max $M or $0." — scored 50/100 and was flagged as just vibes. 19 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 19 claims analyzed: 0 scored under 40, 1 between 40 and 69, and 18 at 70 or above.
Claims analyzed
Coin flip prediction game: $1 per correct guess over M rounds, max $M or $0. — Just Vibes (50/100)
At 0:00
Straightforward setup for a prediction game — classic thought experiment to hook you on probability math.
Why this score: This is pure setup, not a factual claim to verify — just defining a hypothetical game. - Standard coin flip scenario with binary outcomes and linear payoffs. - No predictions or stats yet, so nothing to debunk; it's logically sound as described. - Expecting this leads into Bayesian updating or Kelly criterion next — solid framing.
Original quote: “[0:00] let's consider the following game a coin [0:03] is flipped and remains hidden to you [0:06] your job is to predict what side is [0:08] facing up on the coin if you are correct [0:11] you will earn [0:14] $1 if you are incorrect you will earn [0:18] nothing this repeats for a total of M…”
Designated coin flipper expert always predicts correctly because they know the result — Solid (85/100)
At 0:30
Logically sound setup — flipper knows the outcome, so 100% accuracy as long as they report truthfully.
Why this score: Scenario logic holds within the hypothetical rules.
- Coin flipper sees result before predicting, others don't.
- Assumption of truthful reporting is explicit and reasonable for the thought experiment.
- Distinguishes perfectly informed agent from uninformed guessers.
*Minor nit: real-world 'truthful reporting' isn't guaranteed, but that's outside the game's rules.*
Original quote: “one of the people in the group of experts is designated as the coin flipper at the start of every round they go over to flip the coin cover it up and then rejoin the group each expert then truthfully reports what they predict the coin is showing obviously the designated expert will always be…”
Repeatedly eliminating wrong experts finds the designated one eventually — Solid (85/100)
At 2:30
This is straight out of online learning theory — the halving algorithm converges to the best expert. Solid math.
Why this score: The claim accurately describes an iterative elimination process to identify a 'designated expert' (the consistent predictor), which is a core idea in mistake-bound model of learning theory.
- In each round, incorrect predictors are removed, guaranteeing progress toward the single best expert.
- No fixed round count is a correct caveat — worst-case depends on error rates, but convergence is assured.
- Maps directly to algorithms like the weighted majority algorithm.
Original quote: “remove experts round after round we will eventually reach a single remaining expert who must be the designated expert notice that we have no guarantee on the number of rounds it takes to find the designated expert”
Best case: 1 round if others wrong immediately; worst case many rounds if experts rarely err — Verified (95/100)
At 3:00
Nailed the bounds perfectly — best case instant elimination, worst case logarithmic in errors. Math checks out.
Why this score: Precisely correct analysis of elimination dynamics.
- Best case: All N-1 non-experts err in round 1 → immediate identification (probability-based but possible).
- Worst case: Good experts err rarely → takes ~log(N) rounds via halving, aligning with theoretical bounds in learning algorithms.
- No exaggeration; reflects real variance in expert error rates.
Original quote: “in the best case all of the N minus one other experts predict incorrectly immediately making it take just one round to find the designated expert but maybe the experts are really good and extremely rarely make a mistake in which case it could take many rounds”
Majority vote of remaining experts; eliminates at least half if wrong — Verified (100/100)
At 3:47
Classic majority voting — guarantees halving the pool on mistakes. This is textbook ensemble learning.
Why this score: Core principle of ensemble methods and halving algorithm, spot-on.
- Majority prediction ensures: correct → win; incorrect → ≥50% erred → eliminate them, halving pool size.
- Directly from weighted majority and wisdom of crowds under independence assumptions.
- 'Get the dollar' ties to standard prediction game payoffs; logic is airtight.
Original quote: “we will always predict what the majority of the remaining experts predict if we make the correct prediction then great we get the dollar and a few of our experts disagree but if we don't predict correctly then so must have at least half of the experts so we will sit content with the fact that we…”
Halving algorithm makes at most log n mistakes — Verified (100/100)
At 4:30
Dead on — classic halving algorithm bound in online learning. Math checks out perfectly.
Why this score: This is textbook accurate for the Halving Algorithm in prediction from expert advice.
- Assumes one perfect expert exists; each mistake halves the expert pool.
- Bound of log₂(n) mistakes is standard theoretical result.
- No assumptions violated here; directly from online learning literature.
Original quote: “number of mistakes that we can make by following this policy in other words how many times can we remove at least half of the experts until we are left with only one expert the always correct expert we can compute this upper Bound by removing exactly half of the experts at each step this case is…”
Algorithm works for any binary prediction like stocks or rain — Solid (85/100)
At 5:02
Correct generalization — applies to any binary sequential prediction. Examples spot on.
Why this score: Valid extension of the algorithm's scope.
- Core algorithm is setting-agnostic for binary outcomes (win/loss, up/down, yes/no).
- Examples (stocks, sports, weather) are standard applications in online learning.
- *Minor caveat*: Real-world noise may require weighted version, but claim holds for ideal case.
Original quote: “our algorithm did not depend on the setting of predicting coin flips this means that in any binary decision setting where there are two options we can follow the algorithm for example we could use this strategy to predict if a stock will go up or down or perhaps to predict the winner of a sequence…”
Algorithm needs perfect expert; rarely guaranteed in practice — Verified (100/100)
At 5:34
Nailed the key limitation — perfect expert assumption rarely holds IRL.
Why this score: Precisely correct critique.
- Halving algorithm's log(n) bound requires at least one always-correct expert.
- Real-world scenarios lack perfect predictors, necessitating robust alternatives.
- Sets up natural transition to weighted majority methods.
Original quote: “however critical to this algorithm is the existence and presence of an all-knowing expert of course this is rarely a guarantee instead we should aim to construct a robust prediction system without the guarantee of a perfect predictor”
Halving fails if all experts wrong once; use weights instead — Verified (100/100)
At 6:02
Spot-on failure mode analysis — straight to Weighted Majority fix. Smart.
Why this score: Perfect diagnosis of halving's brittleness and solution.
- If all experts err simultaneously, halving eliminates everyone (classic failure case).
- Weighted trust parameters enable graceful degradation (multiplicative weights).
- This is exactly how Weighted Majority Algorithm was developed.
Original quote: “if all experts make the same incorrect prediction in the first round our algorithm would remove all of them from consideration and we would be back to random guessing so instead of eliminating the experts Al together it makes more sense to maintain a quantity that indicates how much we trust an…”
Trust starts at 1, stays same if correct, halved if wrong — Verified (95/100)
At 6:30
Standard weighted majority setup — checks out perfectly. Classic algo move.
Why this score: This is the exact initialization and update rule from the Weighted Majority Algorithm in machine learning. - Trust (weight) starts at 1 for all experts - Correct predictions: weight unchanged - Wrong: weight *= 0.5 (halved) - Matches textbook description in online learning literature.
Original quote: “[6:26] trust parameter five initialized at the [6:29] value of one between rounds if the [6:32] expert predicts correctly their trust [6:35] value stays the same if they are wrong [6:38] fi gets [6:43] haved”
Mistake when ≥ half total trust predicts wrong — Verified (92/100)
At 7:26
Spot-on definition of when weighted majority errs — no notes.
Why this score: Core prediction rule is accurate: predict the side with higher sum of trust weights. - A mistake occurs precisely when wrong side has ≥50% of total trust F_t - This triggers the halving penalty on the losing (wrong) side's trust - Direct from the algorithm's majority vote mechanism.
Original quote: “[7:24] [Music] [7:41] it remains to determine how many [7:43] mistakes our algorithm makes let F to [7:46] the T denote the total trust at time t a [7:50] mistake is made when at least 1 half of [7:52] f to the T corresponds to the wrong [7:55] prediction”
Halving keeps ≥1/4 correct trust, loses ≥1/4 wrong, total ≥3/4 prior — Solid (88/100)
At 8:21
Math holds tight — bounds are correct for β=1/2 halving parameter.
Why this score: The mistake analysis follows the standard regret bound proof. - Wrong trust (≥F_t/2) halved → loses ≥F_t/4 - Correct trust (≤F_t/2) unchanged → keeps ≥F_t/4? Wait, actually correct keeps whatever it was (could be less), but untouched correct + halved wrong ≥ 3/4 F_t. - Minor phrasing looseness on 'at least 1/4 correct kept' but total trust ≥3/4 F_t is exact. Still solid derivation.
Original quote: “[8:02] which means we maintain a quantity of [8:06] trust that is at least 1/4 of 5 to the T [8:09] but we also lose a quantity of trust [8:12] that is at least 1/4 of f to the T the [8:15] remaining untouched trust corresponds to”
Total trust ≥ (3/4)^T; algorithm makes M mistakes, each multiplies trust by 3/4 — Solid (85/100)
At 8:30
Math setup holds — classic trust aggregation in prediction algorithms. Clean bounds.
Why this score: This establishes the core inequality for total trust in the algorithm. Solid derivation:
- Initial trust sums to *n* across experts
- Each algorithm mistake applies 3/4 factor to *remaining total trust*
- Inductively leads to trust ≤ n*(3/4)^M later
Standard in multiplicative weight updates; no errors here.
Original quote: “summing up what we have it must be at least 34s of 5 to the T but this doesn't yet tell us anything about the number of mistakes that we make for now let us say that our algorithm makes M mistakes and we will try to determine what m is we now know that each time we make a mistake a 34…”
After M mistakes, total trust ≤ n * (3/4)^M — Verified (95/100)
At 9:12
Inductive bound is textbook correct — trust decays geometrically with mistakes.
Why this score: Precisely accurate mathematical induction. Key steps:
- After 1st mistake: trust ≤ (3/4)n
- After 2nd: ≤ (3/4)^2 n
- General: after M mistakes, ≤ n*(3/4)^M (despite minor transcript garble on 'to^ N')
This is the standard upper bound in online learning theory for mistake thresholds.
Original quote: “similarly after the second mistake the amount of remaining trust would be less than or equal to 3/4 2 * n following this pattern inductively after we make the MTH and final mistake our total trust is less than or equal to n * 34 to^ N how else can we bound fight the T”
Best expert's final trust = (1/2)^{best} after best mistakes — Verified (100/100)
At 10:01
Best expert trust halves per mistake — pure math, checks out perfectly.
Why this score: Fundamental and correct. Breakdown:
- Best expert starts at trust 1
- Each mistake: multiplies by 1/2 (halving)
- After *best* mistakes: exactly (1/2)^{best}
This sets up the key comparison for algorithm performance vs. best expert; no flaws.
Original quote: “let's compare the total trust to the trust we have for the best best expert suppose the best expert makes a number of mistakes which we denote as best notice that the best expert starts with trust of 5al 1 and each time it makes a mistake five Habs so one goes to half goes to 1/4 Etc since we have…”
Algorithm mistakes ≤ 2.4 × best expert's + constant — Verified (95/100)
At 10:30
Spot on — that's the exact bound from the classic Weighted Majority Algorithm. No cap.
Why this score: This is precisely correct for the (unrandomized) Weighted Majority Algorithm in prediction with expert advice.
- Standard bound: mistakes ≤ ~2.41 × (best expert's mistakes + log₂(n)), where n is # experts — the '2.4' and 'small additive constant' match textbook results.
- No exaggeration; this is foundational online learning theory from Littlestone & Warmuth (1994).
- Speaker accurately conveys it performs 'similarly' to best expert retrospectively.
Original quote: “the number of mistakes our algorithm makes is at most 2.4 time the best Expert Plus some small additive constant”
Experts rep outcomes; trust = prob; penalty = 1 - εL — Solid (85/100)
At 11:24
Standard extension to multi-outcome via MWU — penalty factor is textbook multiplicative update.
Why this score: Legit adaptation of Multiplicative Weights (Hedge) for multiclass prediction with expert advice.
- Experts per outcome, weights → probabilities: core to randomized forecasters.
- Update rule β = 1 - εL (or ηL): exact for learning rate ε on loss L — common in online convex optimization.
- Minor transcript glitch ('1us') but concept holds; aligns with exp-concave loss handling.
Original quote: “each expert will be a representative for one of the many outcomes the trust value for an expert now corresponds to the probability that we choose that action to predict lastly the penalty factor is 1us Epsilon L where Epsilon is a set parameter and L is the loss”
Advanced algo gets 1× best expert performance (in expectation) — Solid (80/100)
At 12:09
True in expectation for randomized MWU — regret vanishes relative to best, factor →1 asymptotically.
Why this score: Accurate for modern variants like randomized MWU or exponentially weighted forecaster.
- Improves 2.4 to O(1) factor in expectation via randomization and proper tuning (e.g., for log-loss or exp-concave).
- Regret bound: O(√(T log N)) vs best expert's ~0, so ratio →1 as T→∞.
- Caveat: deterministic can't beat ~2; 'in expectation' qualifier is key and correctly stated.
Original quote: “it achieves performance as good as the best expert that is put simply that 2.4 factor from before becomes a one with this algorithm and expectation”
Technique lets non-experts predict almost as well as top experts — Solid (80/100)
At 12:30
Checks out — no-code AI tools and citizen data science make this real. But 'almost as good' needs the full context.
Why this score: The claim aligns with trends in augmented analytics and no-code ML platforms that empower non-experts.
- Tools like these enable predictive modeling without deep expertise, matching expert performance in domains like sports betting.
- Speaker qualifies it well by noting domain examples like basketball.
- Caveat: Real-world results vary by tool quality and data access, but evidence supports the core idea.
Original quote: “[12:30] can perform almost as good as the best [12:32] expert without you having much domain [12:35] expertise that is I might not know a lot [12:38] about basketball but with this technique [12:41] I can predict winners almost as well as [12:44] the best predictor”
See the full analysis with timestamps →