Drawdown Recovery: A Decision Protocol Without Recovery Promises
A drawdown protocol is a decision procedure for losing money, not a promise to recover it. Its first purpose is to prevent the previous balance from becoming a target that dictates the next stake. Its second is to check whether the record, forecasts or execution changed. A losing run can occur under a positive-expectation model, but that observation does not establish that your model still has an edge. The same negative results are compatible with several explanations. Respond with an explicit capital constraint and an investigation, rather than a story chosen from the scoreboard.
Measure drawdown from a consistently defined running peak. Adjust for external deposits and withdrawals or report them separately; otherwise a transfer can create a false recovery or a false loss. Keep free cash, open ticket exposure and pending transfers distinguishable. A peak is not a second balance available to spend. If current capital is below the peak, a cash stake calculated from that old peak represents a larger percentage of the remaining capital. Write down which denominator governs each new position.
A losing streak concerns consecutive losses, while maximum drawdown includes sequences with intervening wins. Neither is determined by the final win count alone. For an explicit independent binary model, a run-length recurrence can calculate the chance of at least one long loss run exactly up to numerical precision. The example below stipulates a constant fifty-four percent win probability for a thousand bets. It is not an estimate of the reader's record. It does not forecast the largest capital loss because stake policy and odds must also be specified.
Calculating loss runs
The recurrence keeps track of paths that have not yet reached the selected loss-run length. A win resets the trailing-loss count; a loss moves the path to the next state. Probability that leaves the final state is absorbed as a hit. This avoids counting overlapping windows as independent events. A single fixed sequence of ten losses has probability q to the tenth power, but a thousand-bet horizon offers many overlapping opportunities. Neither that single-window probability nor a simple binomial total-loss calculation answers the horizon question.
Choose an action rule before a stressful session, while recognizing that its thresholds are policies. One illustrative plan reduces exposure or pauses at a twenty percent drawdown and requires a broader review at thirty-five percent. Those levels are not universally optimal, and no one is required to continue at reduced stakes while investigating. If essential money is involved, a limit is repeatedly overridden or control is deteriorating, the appropriate action is to stop and use protective tools. A smaller wager is not always the right next step.
An audit starts with reconciliation. Include all tickets since the relevant peak, not only losses. Verify accepted odds, settlement, market definitions, model versions and any manual changes. Then inspect forecast calibration, reference prices, roster information, patch applicability and execution. A positive closing-price comparison can coexist with an incorrect forecast or a weak benchmark. A negative comparison can reflect new information or poor execution without proving the entire model obsolete. CLV is one diagnostic; it cannot classify every loss as harmless variance or establish decay from a fixed count of observations.
Audit and loss chasing
Investigate segments only with their sample sizes and selection process visible. A game, market or event may look different because its odds, stake sizes or information sources differ. A group discovered after searching many alternatives needs a later test, not a confident label of confirmed leak. Record missing closing references instead of silently removing those tickets. Where the cause remains unknown, keep it unknown. Paper tracking can continue the investigation without committing more money merely to satisfy a quota of validation bets.
Loss-chasing systems fail a different test. A progression increases required exposure because earlier bets lost, not because the next price or probability improved. The simple doubling arithmetic below shows a funding failure despite money remaining. At odds below two, doubling after losses does not maintain the original profit target either. A sequence can be broken before literal bankruptcy and before a bookmaker's maximum stake is reached. There is no need to invent a universal operator limit or a near-certain ruin percentage to explain that constraint.
Return to a previous stake policy only after a fresh decision about affordable capital and model evidence. A capital threshold can be one condition, but reaching ninety percent of a previous peak does not validate a probability model. Neither fifty nor one hundred favourable price records proves profitability. Document what changed, which evidence was checked, unresolved uncertainty and the new exposure limit. Do not reset the historical peak to hide the remaining drawdown. If you establish a new reporting period, preserve the old series and label the transition.
Recovery arithmetic and decisions
Recovery arithmetic describes required gains, not a timetable. After losing twenty percent, recovering the starting amount requires twenty-five percent on the remainder. Dividing a target gain by expected profit per bet describes a mean-profit horizon under fixed assumptions; it is not the expected first-passage time or a guaranteed number of bets. Paths can recover earlier, later or never within the period. Do not add a calendar deadline or increase volume to force that mean line to arrive sooner.
Write the operational response in a short record: current capital, adjusted peak, open exposure, chosen pause or reduction, audit scope and restart criteria. Use the local tool for the fields it actually exposes, without assuming automatic drawdown alerts or a simulation engine. The protocol is complete when the decision and its evidence are clear. It may end with continued observation, a revised model or a permanent stop. Restoring a past balance is not an obligation that has to be met through more betting.
Worked conditional calculations
Hypothetical independent constant-probability model
p(win)=0.54; q(loss)=0.46; N=1000
state d_j(n): probability of no k-loss run after n bets,
with exactly j trailing losses, j=0,...,k-1
d_0(0)=1; all other states initially zero
d_0(n+1)=p * sum_j d_j(n)
d_j(n+1)=q * d_(j-1)(n), for 1<=j<k
P(at least one k-loss run by N)=1-sum_j d_j(N)
k=7 -> 0.907781631
k=10 -> 0.203677582
This is not a binomial total-loss count or a drawdown probability.
Illustrative capital arithmetic, no external flows
peak H=10000; current B=8000
drawdown=(H-B)/H=0.20
required gain from B back to H=(H-B)/B=0.25
stake 100 based on old peak -> 100/8000=1.25% of current B
stake 80 -> 1% of current B
stake 40 -> 0.5% of current B
These amounts compare policies, not survival guarantees.
Hypothetical fixed-price forecast, not a recovery promise
p=0.54; odds=1.95; per-unit EV=0.54*1.95-1=0.053
stake=0.5 original units; mean profit per bet=0.0265 units
target gain=11 original units
11/0.0265=415.094339623 bets on a mean-profit line
not the expected first hitting time and not a guaranteed recovery
break-even p at 1.95=1/1.95=0.512820513
at p=0.513, EV=0.00035, not exactly zero
Doubling example, starting capital 100, initial stake 1
six consecutive losses: 1+2+4+8+16+32=63
remaining=37; next requested stake=64, not fundable
after k losses and then a win at 1.95 on stake 2^k:
net=0.95*2^k-(2^k-1)=1-0.05*2^k
k=5 -> net=-0.6, despite a final win
loss count does not improve the next event's conditional price.
