China's sports lottery is not a betting market — it's a formula

I measured the margin on 897 matches. The surprising part is not the size of the fee, but that
the fee barely moves — and what that implies.

---

The Chinese Sports Lottery (竞彩) publishes fixed odds on football matches. I wanted to know what
they actually cost, so I wrote down the three-way odds for 897 matches between 2026-05-17 and
2026-09-25, and paired each of them with the same fixture's "99 bookmakers average" closing
odds from the European market.

The measure is the overround: the sum of the implied probabilities.

overround = 1/home + 1/draw + 1/away
payout    = 1 / overround

If the odds were fair, overround would be exactly 1.0. Real books are always above 1.0, and the
excess is the margin — the fee you pay before your opinion about the match is even relevant.

Here is what 897 matches gave me:

market	n	overround	margin	σ	range
Chinese lottery, 1X2	897	1.1292	12.92%	0.0006	1.1251 – 1.1299
Chinese lottery, handicap	897	1.1291	12.91%	0.0006	1.1251 – 1.1299
European average (99 books)	897	1.0685	6.85%	0.0109	1.0403 – 1.1085

The lottery is more expensive in 897 of 897 matches. No exceptions. The mean gap is
6.07 percentage points.

The σ is the actual finding

Everyone can guess that a state lottery is expensive. The interesting column is the fourth one.

Across 897 matches — different leagues, different days, different kick-off times — the lottery's
overround stays inside a band 0.0048 wide. In percentage terms, the margin never leaves
12.51%–12.99%. The European average swings 19× more, because it is set by people: it follows
league, liquidity, and how close the match is to kick-off.

That difference is not a curiosity. It tells you what kind of object you are looking at:

• A market price is produced by the interaction of many participants. It moves, because it aggregates disagreement. A price that moves can be wrong — and that is exactly what professional syndicates spend millions looking for.
• A formula price is produced by a fixed payout parameter. It does not move. A price that does not move cannot be wrong.

The confirmation is inside the same lottery: two different play types — 1X2 and handicap — land
on 12.92% and 12.91%. The same parameter applied to everything. That is what a formula
looks like.

The practical upshot: there is no mispricing to find here. Not "hard to find" — nothing to find.
You are not looking for a bad price in an efficient market; you are looking at a posted fee.

What the fee does to multi-leg bets

The margin compounds per leg. Per 100 staked, long-run expected return:

legs	returned	taken
1	88.6	11.4
2	78.4	21.6
4	61.5	38.5
6	48.2	51.8
8	37.8	62.2

A 6-leg accumulator gives up more than half. For scale: American roulette with double zero
takes 5.26%. A 6-leg ticket here takes roughly 10× that.

This is also why the usual advice — "find value, bet singles, avoid parlays" — is right about
parlays and hopeless about value. You would need a 12.92% edge to break even. Professional
betting teams build models to find 2–4%. The bar here is 3–6× higher than the edge that
professionals consider a good year.

I also tested this on myself

To be fair to the other side, I ran 15 betting rules over a 1,617-bet sample — favourite
backing, home/draw/away, handicap variants, odds-band filters, a model-value filter, and a
daily 2-leg combination. Stated precisely:

• 12 of 15 rules are negative; 9 are significantly negative (95% CI upper bound below 0). Worst: −58%.
• 3 of 15 are positive: +23.78% (n=37), +5.79% (n=127), +3.94% (n=52) — and all three have 95% confidence intervals that cross zero (±116%, ±15%, ±86%). The largest makes +143% and +43% in its first two months and −100% in each of the last three.
• 0 of 15 pass the rule I fixed before looking at the data: n ≥ 300 and CI lower bound > 0.

So the honest summary is not "every strategy loses." It is: nothing survives the sample-size
bar, and the positive results are indistinguishable from noise.

One more result worth keeping, because it is the one thing that is real: hit rate rises
monotonically with confidence tier, from 46.6% to 82.2%. The information is there. But the average
odds fall to 1.23, so breaking even needs 81.3% — and the best tier hits 82.2%. It lands exactly
on the margin line. That is what a 12.92% fee looks like from the inside: you can be right, and
still be paying tolls.

Limitations

• One market, one period: a four-month season fragment, not a full year.
• The benchmark is the 99-bookmaker average close, not a single book's live price.
• Overround measures the price of entry, not any individual's loss rate — that depends on what and how much you actually stake.
• The σ comparison is well supported (overround is a per-match constant). The finer breakdowns — by league, by month — are not, and are not claimed here.

Reproduce it

Everything ships as raw data plus scripts, standard library only, no API keys:

git clone https://github.com/quietnumbers/china-sports-lottery-overround
cd china-sports-lottery-overround
./reproduce.sh

overround.py --check recomputes the headline numbers from the raw odds CSV and exits
non-zero if any of them moved. backtest.py recomputes the 15-rule table from the per-bet
files. The dataset is also published as a GitHub release if you would rather grab the CSVs in one file.

Disclaimer

This is data analysis, not betting advice. I do not recommend matches, sell selections, or
promise any return. I measured this number precisely to make the opposite point: in this market
every strategy must first clear a 12.92% wall. If it cannot, you are paying someone else's toll.