What free betting tips actually return

Do free tips beat the closing line?

Closing line value is the only measure of a tip that converges before the results do. If a selection is reliably available at a bigger price when it is published than when the market closes, the tipster has found something the market had not priced — win or lose. A win rate needs hundreds of picks to mean anything; a beat rate needs far fewer.

Beat the close

42.9%

95% CI [41.3%, 44.4%]

1,689 of 3,939 paired quotes · null is 50%

Mean line movement

+0.30%

Entry price against close, clamped at ±30%

The interval sits entirely below 50%. These tips are, on average, published at a worse price than the market closes at. The sample is large enough that this is not noise: by the time a free tip is public, the move it was pointing at has usually already happened.

What they returned

Only picks published with a real price can be turned into a return. Where a source publishes a probability instead, a price can be derived from it — but a derived price carries no bookmaker margin, so the profit it implies beats anything you could have placed. The two are never averaged. Everything below is advertised prices only.

Priced picks

21,058

Published with a real price

Units staked

115,392

Flat stakes

Profit

+9556.9

Units

Win rate

54.9%

95% CI [54.3%, 55.5%]

26,424 graded, void excluded

Yield across the priced corpus is +8.28%. Read it against §1: a portfolio can show a positive yield over a few thousand picks while having no measurable timing edge, and when those two disagree the beat rate is the one that keeps predicting.

By source

Every source tracked, its entire settled output, no cherry-picking of good months. Yield appears only where the source publishes real prices; the beat rate only where at least 20 picks carry both an entry and a closing quote.

How this was measured

1. 01Tips are captured when they are published, before the result exists. Nothing is added to the corpus retrospectively, so a losing pick cannot quietly fail to arrive.
2. 02Grading is against the final result of the fixture, by the same rules for every source.
3. 03Void picks are settled but put no stake at risk. They count in “settled” and not in the win rate.
4. 04The entry price is the market quote at the time the pick was first seen, not necessarily the price the tipster themselves got. Closing is the last pre-kickoff quote. Both come from the same feed, so the comparison is like for like.
5. 05Line moves beyond ±30% are clamped. A move that size is team news or a capture error, and either way it is not the tipster’s timing.
6. 06Intervals are 95% Wilson score intervals, which stay inside [0,1] at small samples where the normal approximation does not. No interval is reported below 10 graded picks, and no beat rate below 20 paired quotes.
7. 07A missing stake is treated as one unit — level staking, which is what a source publishing no stake is doing.

Full grading rules are in the methodology.

Independence and availability

Tipzy is not affiliated with any source named on this page. We track them; we do not represent them, and they have no say in what these numbers say. Where a figure is unflattering to a source, it is printed unflattering.

Per-tipster detail, including every individual pick, is on the records. A tipster who believes their record here is wrong can claim it and say so.

Computed

2026-09-02 13:47 UTC

Cadence

Every request, from the live corpus

Revisions

None — the corpus only grows

By line, metric and period

Table 1 counts every Over/Under pick as one bucket, which quietly averages Over 2.5 with Over 3.5 — two different bets sold at two different prices. The table below splits the corpus by the exact line instead, so each price stands on its own. Markets that carry no line at all — match winner, both teams to score — are not dropped; they are grouped under "No line" below, because a bucket that silently disappears reads as no data rather than as not applicable.

Win rate is not the column to compare across rows here: a shorter line wins more often and is priced shorter for exactly that reason, so a higher win rate on Over 1.5 than Over 3.5 says nothing about which was the better bet. Yield is what to read — it already prices in the difference. As with every yield figure on this page, the three tables below are advertised prices only: a price a source derived from its own claimed probability carries no bookmaker margin, and must never blend into a Settled count, win rate or Yield next to one that does.

The same corpus split two other ways: by what is being counted — goals, corners, cards, shots — and by which part of the match the bet covers.

What sources claim, against what we measured

Every observation we capture also records the ROI a source was advertising for that tipster at the moment we scraped the tip, and the tip count the source made that claim over. Of the 14 sources this site tracks, only Oddspedia publishes a claim we have captured. The table below sets that claim against our own measured yield for the tipsters Oddspedia covers.

The two numbers are not comparable like-for-like. A source's claimed ROI covers a tipster's entire published history -- including tips we never captured, and periods before Tipzy existed. Ours covers only our own observation window, priced at the prices the source itself advertised. A claim made over several thousand tips next to a settled sample in the dozens is not evidence that either number is wrong; it is a difference in what each one was measured over, which is why both sample sizes are printed on every row. Only tipsters with at least 30 graded picks appear here -- below that our own number is not solid enough to set anyone's claim against.

Oddspedia is the only source with a claim in this table; its median gap across the 1 tipster above is +47.3pp.

Consensus against the crowd

When several tipsters back the same fixture and market, how often does the most-backed selection land, and what does each side actually return? Two tips are the same fixture when they share an event, or when they name the same two teams on the same UTC day. Consensus is measured per fixture and market: three tipsters on Over 2.5 and one on the home win are not a four-way disagreement, they are betting different things, so each market on a fixture is its own group. A group counts only with two or more tips sharing a market and a resolvable selection, and only when one selection strictly beats every other individually — a tied top spot has no winner to name.

That top selection is the most-backed, not necessarily a majority: on a market with three or more selections it only has to beat each rival on its own, not outnumber them combined. A WINNER market split 5 Home / 4 Draw / 3 Away below reads as 5 against 7 — Home is the most-backed pick, but the other two selections together have more tips than it does. The "rest" side is a mixture of whichever different bets those tipsters actually made, not one shared opposing view.

A win rate comparison here would be close to tautological: the side more tipsters back is usually the favourite, and favourites win more often at shorter prices — that is what a shorter price means, not evidence the crowd called it right (see the homepage's strike-rate/yield pair for the general case). Yield is what actually says whether either side paid, so it leads below; win rate stays as context underneath it, over the same population. Both are measured over advertised prices only, the same rule §2 states for the whole corpus — a price a source derived from its own claimed probability carries no bookmaker margin, so it is excluded from this section entirely, not just from yield.

This section speaks for 55.7% of the settled rows we hold — 16,000 of 28,718 rows carry a fixture identity we can group on. The rest carry neither an event nor a verified kickoff and are excluded rather than guessed at.

These groups are backed by 668 distinct tipsters across 13 distinct sources.

Most-backed yield

+8.61%

3,561 advertised picks · mean price 2.01

Rest yield

-22.22%

523 advertised picks · mean price 2.44

Most-backed win rate

55.1%

95% CI [53.4%, 56.7%]

3,561 graded, advertised only

Rest win rate

49.9%

95% CI [45.6%, 54.2%]

523 graded, advertised only

The most-backed selection yields +8.6% over 3561 priced picks; the rest of the field yields -22.2% over 523 priced picks. The most-backed selection paid better here. The most-backed selection is priced at a mean of 2.01, the rest at a mean of 2.44 -- sides priced this differently are expected to post different strike rates on their own, before either one has actually called anything better.

By odds band

Betting markets are widely reported to show a favourite-longshot bias: short prices pay closer to their true odds than long ones do, because longshots attract more casual money than their real chance of winning justifies. If this corpus shows that pattern, yield should fall as price lengthens — not win rate, which falls as price rises by construction and says nothing about whether the price was fair. The table below splits the corpus into five price bands and reports both, so the two can be told apart.

This section speaks for 81.1% of the settled rows we hold — 22,411 of 28,718 rows carry an advertised price. As with every yield figure on this page, prices derived from a source's own claimed probability are excluded entirely, not just from yield — they carry no bookmaker margin and would inflate every band's return above anything placeable.

[1.01, 1.50) returns -2.5% against 5.00+'s +46.9% -- a 49.4pp gap, within what ordinary variance between two samples this size would produce. This corpus does not show favourite-longshot bias at the current sample.

By staked conviction

When a tipster stakes more than usual, do they do better? Stake units are not comparable across sources — a Blogabet 5 (out of 10) and a different source's 5 do not mean the same thing — so raw stake size is never compared across the corpus. Instead, every priced pick is measured against the tipster who made it: below, at, or above that tipster's own median stake. A tipster whose bigger bets have done worse than their smaller ones is exactly as much a finding here as the reverse.

70.3% of the settled rows we hold publishes a stake at all — the rest are grouped under "No stake published" below rather than assumed to be level stakes. Each tipster's median is taken over every staked pick they made, any odds kind; only the yield figure below is restricted to advertised prices, the same rule §2 states for the whole corpus.

A tipster's own median needs a real sample to mean anything: with only a handful of staked picks, sorting always produces some below and some above the median regardless of what the tipster actually did. Tipsters with fewer than 10 staked picks are withheld from below/at/above entirely and grouped under "Too few staked picks to rate" instead — 398 tipsters, 1,399 staked picks in total, over the whole settled corpus.

"At own median" is not a third conviction level to read alongside below and above: it also catches every tipster who stakes identically on every pick, since a flat stake is, by definition, always at its own median. A source-wide convention of flat staking can make this the biggest row in the table without saying anything about conviction at all.

Above-median stakes return +13.3% against below-median's +0.7%, a 12.6pp gap that is bigger than ordinary variance between two samples this size. When these tipsters bet bigger, they have done better.

By publishing lead time

Does a tip published well ahead of kickoff do better than one published minutes before? Lead time is measured as feed kickoff time minus publish time, and gated on feedStartDate only — sourceEventDate and llmEventDate are display fallbacks that were never feed-verified, and the LLM date backfill calibrated at 0% accuracy within ±1 day, so a lead time computed off either would be fiction, not a measurement. A tip whose publish time lands after its own feed kickoff — a historical archive scrape, or a wrong timestamp — is not a lead time at all; it gets its own labelled bucket below rather than being folded into a real one or dropped.

This section speaks for 34.0% of the settled rows we hold — only rows carrying both a publish time and a verified feed kickoff can be placed. As with every yield figure on this page, the table below is advertised prices only.

2 rows across the whole settled corpus — not just the advertised-priced ones the table below is restricted to — come out with a negative interval, publish time after feed kickoff. None are dropped: the table's own "Negative (published after kickoff)" row is the advertised-only slice of that same 2, so it can read smaller than this count.

48h+ returns +29.1% against Under 2h's -3.9%, a 32.9pp gap that is bigger than ordinary variance between two samples this size. A longer lead time has gone with a better return here.

By team

Which teams show up most often in the fixtures this corpus tips, and how do the tips on those fixtures settle? This is not the same question as which teams tipsters back most: homeId/awayId say which teams a fixture involved, not which side a tipster picked — that is the selection, and it is not always a team at all (an over/under tip backs neither side). A tip on Arsenal vs Chelsea is counted once for Arsenal and once for Chelsea, so the counts below sum to roughly twice the number of settled rows — they are not a partition of the corpus, and no two rows in the table below are mutually exclusive.

This section speaks for 96.3% of the settled rows we hold — 27,669 of 28,718 rows carry at least one participant id. 1,049 rows carry neither and are excluded rather than guessed at. As with every yield figure on this page, the table below is advertised prices only: 6,775 distinct teams appear in an advertised-priced tip, and it shows the top 20 by sample size.

Fixtures involving the most-tipped teams yield -8.7% over 1500 priced picks, against +8.3% across the whole corpus. A 17.0pp gap against the popular teams, bigger than ordinary variance would produce.

Advertised price against the feed's own quote

When a source advertises a price, how close is it to what the market was actually quoting? For every ADVERTISED pick carrying a captured feed snapshot, the gap is the advertised price against that snapshot — clamped at ±30%, the same clamp and the same reason §1 applies to closing-line movement: a bigger divergence is a capture bug, not a quoting practice, and it must not drag a source's mean.

oddsSnapshot is the first feed quote we captured, taken at firstObservedAt -- not necessarily the moment the source published its tip. If a source published its tip hours before we first saw it, the two prices are not contemporaneous, and part of the gap below is staleness between two quotes taken at different times, not evidence the source quoted a price nobody could get.

This section speaks for 3.0% of the settled rows we hold — only ADVERTISED rows carrying a captured feed snapshot can be placed. Sources below 20 such rows are withheld entirely rather than shown with a thin sample.