How much of my bankroll should one tennis bet be?

A common range is 1-2% of the bankroll per bet, kept flat regardless of how confident the pick feels. The test for any figure is a drawdown test rather than a return target: our own record contains a 9.91-unit peak-to-trough fall, which costs about 10% of a bankroll at a 1% unit and about 50% at a 5% unit. Pick the size where the worst run you can plausibly hit still leaves you able to keep betting the same way.

How many losing bets in a row is normal?

Longer than most people expect. At a 60% hit rate, a five-loss streak has roughly a 1-in-100 chance of starting at any given bet, which means a sample of a few hundred bets is expected to contain about two of them. In our 221 settled picks the longest run was five straight losses over three days in July 2026. A streak that size is information about variance, not about whether the method works.

Is flat staking better than the Kelly criterion?

It depends entirely on how accurate your edge estimates are. Kelly is optimal when the estimated edge is correct and aggressive when it is overstated, and overstated edges are the common case - in our own log, the price band holding the largest claimed edges returned 50.0% winners from 42 picks against 62.8% from the 156 picks at shorter prices. Flat staking, or a fraction of Kelly, is the sensible response to estimates you cannot fully trust.

Should I increase my stake after a losing run to recover the loss?

That pattern turns a normal losing stretch into a bankroll-ending one, because the streak has no memory of what you have already lost and the larger stakes arrive exactly when the run is longest. Our own worst stretch took twelve days to bottom out and another twelve to recover at constant stakes; the same stretch with escalating stakes does not recover at all. Deciding the stake before the run starts is the entire point of a staking plan.

Bankroll Management for Tennis Betting: Size for the Drawdown, Not the Average

Bankroll management is not a money question, it is a variance question: how large can one stake be so that the worst stretch your method will produce does not end your ability to keep applying it? The average tells you nothing useful here. Our own tracked record as of 22 September 2026 reads 262 picks, 221 settled, 133 wins, 88 losses, a 60.2% hit rate, +22.44 units and a +10.2% ROI (headline-stats.json). That is the average. The path underneath it is the part that decides your stake size: between 29 June and 11 July 2026 the running total fell from +7.90 units to -2.01 units, a drawdown of 9.91 units, and did not get back to its old high until 23 July. Nothing about the method changed in those twelve days.

A unit is a percentage, not an amount

A unit is the fixed share of your bankroll that one standard bet costs. It is expressed as a percentage precisely so that it shrinks when you are losing and grows when you are winning, without any decision on your part. One unit at 1% of a 2,000 bankroll is 20; after a 10% drawdown the same rule stakes 18.

Two consequences follow immediately, and both are more important than the exact percentage you pick.

First, a unit only means something if every bet is the same size. The moment stake size varies with how confident you feel, your record stops being measurable: a 60% hit rate is meaningless if the 40% you lost happened to carry double stakes.

Second, unit size and price are not the same lever. Across our 221 settled picks the average price taken was 1.84 and the median 1.88 (picks.json, generated 22 September 2026) - short enough that a losing bet costs more than a winning one returns. If you are not sure what a price of 1.84 claims about a player’s chances, the implied probability calculation is the first thing to get comfortable with, because every sizing rule below assumes you can read a price as a probability.

Size for the drawdown you have not had yet

The useful exercise is not “what do I expect to make” but “what does the worst run in my own log cost me at this stake size”.

From our log: the longest losing streak was five consecutive settled losses, between 2 and 4 July 2026. The worst peak-to-trough fall was the 9.91 units described above, and the running total spent four days below zero in the first month of tracking. Translate that into bankroll terms:

unit = 1% of bankroll  ->  9.91 units = -9.9% of bankroll
unit = 3% of bankroll  ->  9.91 units = -29.7% of bankroll
unit = 5% of bankroll  ->  9.91 units = -49.6% of bankroll

The first line is an unpleasant month. The third line is a bankroll that needs to roughly double just to get back to where it started, and it was produced by a stretch of results that a positive, working record already contains.

And 9.91 units is not the worst case. It is the worst case in three months of one particular sample. A longer record will contain a deeper hole, for the same reason that a longer coin-flipping session contains a longer run of tails. Sizing off the deepest drawdown you have personally seen is already optimistic.

Why a 60% hit rate still produces losing runs

A five-loss streak looks like something broke. It is not. At the observed 60.2% hit rate, the chance that any given settled pick begins a run of five straight losses is roughly 0.398^5, about 1 in 100. Over 221 settled picks there are about 217 places such a run could start, so the expected number of five-loss runs in a sample this size is a little over two. We had one. Finding fewer than expected is not evidence of anything either.

The same arithmetic applied to the bottom line is more sobering. Treating the observed 60.2% as if it were the true rate and flat-staking one unit at the average price of 1.84, the expected return over 221 bets is roughly +24 units with a standard deviation of roughly 13 units. A 95% band around that runs from about -3 to about +50 units. Our actual +22.44 sits comfortably inside it - and so would a result of zero.

Two caveats belong with that calculation rather than in a footnote. It uses the hit rate measured on the same picks it is evaluating, which is circular; and it collapses a range of prices into one average, which understates the spread. It is an order-of-magnitude check on how wide the range of plausible outcomes is, not a forecast. The point survives both caveats: over three months, sizing decides where in that range you can still afford to be standing.

Flat stakes, Kelly, and why the gap between them is an estimation problem

Flat staking means every bet is one unit, regardless of perceived edge. Kelly staking sizes each bet in proportion to the edge you believe you have, which is mathematically optimal if - and only if - your edge estimates are accurate. That condition is doing enormous work.

Our own data gives a reason to distrust large claimed edges specifically. In the 2.00-2.50 price band, where the bigger claimed edges tend to sit, 42 settled picks produced exactly 50.0% winners and +1.09 units, against 62.8% and +23.60 units from the 156 picks priced between 1.60 and 2.00. Those sub-samples are small and the split was found after the fact, so it is a flag rather than a finding - but it points the same direction as the broader pattern described in favourite-longshot bias, where the largest apparent edges are the least reliable ones.

Full Kelly converts exactly that unreliability into oversized stakes. The practical middle ground - a fraction of Kelly, or flat stakes with a hard cap - is compared in detail in our note on flat staking versus the Kelly criterion. The short version: if your edge estimates are noisy, sizing that is insensitive to them is not a compromise, it is the correct response to the noise.

What bankroll management cannot do for you

Sizing controls how long you survive a negative expectation. It does not turn one positive. If the underlying method has no edge, careful staking changes the speed of the outcome and nothing else, which is the uncomfortable half of the argument in is value betting actually profitable.

It is also worth being precise about what our own record does and does not establish. The closing-line measurement - whether we beat the market’s final price, independent of results - currently covers 69 of 262 picks, with a mean of +2.08 percentage points and a 95% confidence interval from -1.45 to +5.61. That interval includes zero. The number is encouraging and it is not yet evidence; we publish it, and every loss behind it, for the reasons set out in why we publish our losses.

So the honest sequencing is: get the method right first, then size it so a normal bad run stays survivable. A useful default is a unit of 1-2% of a bankroll you have consciously separated from money you need, a fixed stake regardless of confidence, and a rule written down in advance for what happens after a losing week - because the decision made during one is reliably worse.

Betting is for adults aged 18 or over. Stake only what you can afford to lose, and if the size of a stake is affecting decisions you make away from betting, that is the signal to stop and seek support rather than to adjust the staking plan.