Plinko Strategy Results: Compare All Three Recorded Samples

The 3,000 actual drops and how replays work

The raw material is 3,000 actual paid demo drops on a 16-row Plinko board, played in FUN credits on a public international demo between 09:16 and 09:28 UTC on 29 October 2026, with mean latency around 140 ms per drop. There were 1,000 drops at low risk, 1,000 at medium and 1,000 at high, each at a flat 1 FUN stake. Low risk returned 996.1 FUN (net −3.9, observed RTP 99.61%), medium returned 1,004.8 FUN (net +4.8, observed RTP 100.48%), and high returned 846 FUN (net −154, observed RTP 84.6%). Gross return means everything the board paid back; net subtracts the stakes. The medium run’s gross 1,004.8 FUN is not profit — the +4.8 FUN net is.

Every strategy below is a model replay over this fixed recorded history, not new drops. A replay asks: given these exact outcomes in this exact order, what would a given staking rule have done to a 1,000 FUN bank? That distinction matters because a replay cannot change the physics; it only changes exposure per drop. You can reproduce the board behaviour on our Plinko simulator, tune the same knobs on the settings page, and check the pocket maths on the odds breakdown.

The theoretical return of each board is just under 99% (low 98.999%, medium 98.988%, high 98.976%), so the long-run expectation of every replay is negative. Observed results above or below that line are sampling noise around a negative mean, and nothing here is a claim of reliable or proven future profit.

• Actual data: 3,000 paid demo drops, 1 FUN each, all in FUN credits.
• Replays: staking rules applied to that frozen history; no new randomness.
• Gross return = total paid back; net = gross minus turnover.
• Completed 1,000-drop runs are reported separately from early stops.

Reading the table, graphs and holdout grid in plain language

The rendered results table, the bank-line graphs and the 216-grid holdout display appear separately on the page; here is how to read them without a statistics background. Each table row is one strategy at one risk level. 'Turnover' is the total FUN wagered across the run; 'net' is what the bank actually gained or lost; 'max drawdown' is the worst peak-to-trough dip the bank suffered mid-run. A concrete case: the medium fixed-stake line shows turnover 1,000, returned 1,004.8, net +4.8, drawdown 28.2. That means the bank was never more than 28.2 FUN below its last high point, and finished 4.8 FUN ahead.

The bank-line graphs plot the bank after every drop. Arithmetic example from the percentage plan: starting bank 1,000 FUN, the first stake is floored to 1% of bank = 10 FUN. A 0.3x partial return pays 3 FUN, so the bank falls to 993 FUN, and the next stake is 1% of 993 = 9.93, floored to 9 FUN. Note that a 0.3x return is always a net loss on that drop — you staked 10 and got 3 back — even though money 'came back'. The graphs make these slow bleeds visible as gentle downward slopes, while high-risk replays show jagged cliffs.

The 216-grid holdout display shows one dot per parameter set tested on the first 700 drops per risk, with the chosen set highlighted and then re-scored on the final 300 drops per risk that the search never touched. Reading it is simple: dots are training scores, the highlighted dot’s second score is the honest one. If a dot looks brilliant in training but collapses in holdout, it was curve-fit noise.

• Turnover ≠ loss: high turnover with small net loss means heavy recycling of the same bank.
• Drawdown is measured intra-run, not just at the end.
• Holdout dots start a fresh 1,000 FUN bank and an empty pattern history.

Fixed stake and 1% of bank: the completed baselines

Fixed 1 FUN stakes completed all 1,000 drops at every risk. Low lost 3.9 FUN, medium gained 4.8 FUN, high lost 154 FUN. The medium result is the headline positive full-history replay, and its 100-drop windows show how luck clustered: windows three (+9.2), eight (+10.3) and ten (+9.0) did all the work while five other windows were negative. The high-risk fixed run never had a single break-even drop — 200 profitable drops against 800 losers — because the high board pays either a multiplier above 1x or a 0.2x partial, and 0.2x is always a net loss.

The 1%-of-bank plan also completed all 1,000 drops everywhere, but it multiplied turnover roughly nine-fold (about 9,009 FUN at low risk) and converted the same drop outcomes into bigger absolute losses: −36 FUN low, −11 FUN medium, and a severe −790.2 FUN at high risk. At high risk the bank ended at 209.8 FUN after an 893.2 FUN drawdown. The mechanism is pure arithmetic: percentage staking raises stakes after wins and lowers them after losses, but with a negative-expectation board, more turnover means more edge paid. The table’s 'edge cost of realised turnover' column shows this directly — about −90.2 FUN of theoretical edge on the low-risk 9,009 FUN turnover.

Small banks amplify these effects. With 1,000 FUN and 1% stakes, a single high-risk stake is only 10 FUN, yet the plan still lost 79% of the bank because losses compound when stakes stay proportionally large during downswings. With a smaller hypothetical bank, the whole-FUN flooring rule bites sooner and distorts the intended fraction.

• Fixed stake: the only plan where all three risks finished with tiny absolute nets relative to bank.
• Percentage: completed everywhere but turned small edges into large absolute losses via turnover.
• 0.3x and 0.2x partial returns are always net losses on the drop.

Martingale family: caps, resets and bank failure

Classic Martingale (double after a net loss) completed 1,000 drops at low risk with a small −3.4 FUN net and a maximum stake of 32 FUN, because low-risk losing streaks never grew long enough to require a 64 FUN stake. At medium and high it died early: the medium run hit the stake cap after only 13 drops (−72.6 FUN net), and the high run after 35 drops (−27.6 FUN). These are early-stop results and must not be compared with completed 1,000-drop runs as if equivalent — they simply ran out of room to double.

The capped Martingale uses stakes 1, 2, 4, 8, 16 and then resets after the fifth net loss; a profitable result resets the ladder, and a break-even result holds the current step. This completed all 1,000 drops at low (−9.6 FUN, one reset) and medium (−32.5 FUN, 51 resets). But at high risk it destroyed the bank: 907 drops, 91 resets, final bank 3.8 FUN, stopped because the bank could not fund the next stake. Cap versus bank failure is the key distinction: a stake cap ends the run with money left (the medium Martingale stopped with 927.4 FUN), while bank failure means the rule itself consumed everything.

D’Alembert (add 1 FUN after a loss, subtract 1 after a win) was gentle at low risk (−1.6 FUN, max stake 6) but catastrophic at medium and high. Stakes ballooned to 89 and 60 FUN, and the runs stopped for insufficient funds at drops 233 and 110 with losses of −937.5 and −950.4 FUN. Fibonacci behaved similarly: fine at low (−2.9 FUN), capped out at medium after 63 drops (−64.5) and at high after 50 drops (−128.6). The lesson is mechanical: any progression that raises stakes during losing streaks collides with high-risk streaks of 13, 15 or 30 losses.

• Martingale early stops are stake-cap events, not bank failures — the bank survived.
• Capped Martingale’s high-risk ending (3.8 FUN) is a genuine bank failure.
• D’Alembert and Fibonacci fail slower than Martingale but for the same reason.

Paroli, 1–3–2–6, and stop/restart rules

Paroli (double after wins, up to three steps) completed all 1,000 drops at every risk: −10.9 FUN low, +4.0 FUN medium, −186.2 FUN high. The medium Paroli result is the second positive full-history replay in the study, and it makes sense mechanically: Paroli concentrates bigger stakes on winning streaks, and the medium history happened to contain profitable clusters. The 1–3–2–6 sequence also completed everywhere (−3.3 low, −33.8 medium, −163 high), recycling more turnover (2,454 FUN at low) for no benefit.

The fixed stop rule (stop at −20 or +20, evaluated strictly after each drop settles) changed nothing at low and medium risk, because those banks never moved ±20 FUN in a single settled step early on. Both runs are identical to fixed stakes. At high risk it stopped after 101 drops at −20.4 FUN. The restart variant keeps the bank and continues after each stop with no time gap and no free observations. At high risk it restarted 14 times and ended at the same −154 FUN as plain fixed stakes, because restarting does not change the expectation of the drops that follow.

Stop/restart systems reshape when you lose, not whether. A stop rule truncated the high-risk loss at −20.4 FUN by quitting early — a real, if accidental, saving of 133.6 FUN versus riding all 1,000 drops — but the restart version gave it all back by continuing.

• Paroli medium: +4.0 FUN, the second positive completed replay.
• Stops are evaluated after settlement; restart keeps the bank and re-enters immediately.
• Stopping early 'saved' money at high risk only by reducing exposure.

Frozen pocket-pattern boosters on fresh data

A100 reads the previous five settled pocket distances from the centre. It requires, in order: 1–2 pockets away, at least 3 away, 1–2 away, 1–2 away, at least 3 away. A centre hit does not satisfy either category. The next stake is 100 FUN; the next eight drops use the ordinary 1 FUN stake before another trigger is allowed.

B50 and B100 use the same three-drop trigger: the earlier pocket must be at least 4 places from the centre, followed by two pockets exactly 1 place from the centre. They stake 50 and 100 FUN respectively on the following drop, then use three ordinary paid drops as cooldown. C25 instead checks the previous 16 pockets: if no more than one was the exact centre, it stakes 25 FUN on the next drop and then uses a three-drop cooldown. Every observation is paid; previously boosted drops are still part of the recorded history.

The wider search varies boost and cooldown, so B50 and B100 become the same trigger family with different inputs. We count that family once: 3 distinct trigger families × 9 boost amounts × 8 cooldown values = 216 unique configurations. Four named frozen presets remain in the baseline comparison. This avoids counting the same behaviour twice.

A concrete example from this site’s Normal sample: drop 24 followed pockets 12, 7, 7 (indices 0–16), whose centre distances were 4, 1, 1. B100 therefore selected a 100 FUN stake using only those earlier results. Drop 24 then landed in pocket 11, paying 1.5×: 150 FUN returned and +50 FUN net. The outcome was not known when the stake was chosen. This single trace explains the trigger; it is not evidence that the pattern predicts a later drop.

Results: A100 lost at low (−122.7) and medium (−104.1) but gained +123.2 FUN at high risk, where eight boosted drops met a history containing a 26x hit. B50 lost everywhere (−8.8, −259.8, −271.6), and doubling the boost to B100 roughly doubled the damage (−13.8, −529.8, −391.6). C25 was the star: +0.9 FUN low, +139.2 FUN medium (29 boosted drops), −278.8 FUN high.

Exposure explains everything. Boosted drops are just bigger bets on the same negative-edge board, so a pattern 'works' when its boosted drops happen to land well. C25's medium gain came from 29 boosted drops during a friendly stretch; the same rule at high risk met the brutal 800-loss history and bled. Near-edge pockets (the 0.5x/1x region) hit constantly and drive day-to-day results, while the rare tail (110x, 1000x) has about a 3% chance of appearing even once in 1,000 drops — no pattern rule can summon it.

• A100 high: +123.2 FUN — a positive completed replay driven by 8 boosted drops.
• C25 medium: +139.2 FUN — the largest positive net in the whole study.
• Boosts raise exposure; they do not change pocket probabilities.

The 216-parameter search and why the offline holdout matters

Finally we searched 216 parameter sets across the three distinct pattern families (base 1, varying boost and cooldown). Each was trained on the first 700 drops per risk with the objective of maximising the worst net across three training segments, requiring completion, and breaking ties toward lower boost then longer cooldown. The selected configuration — boost 15, cooldown 8 — was chosen on history only, before any holdout was examined. Its training nets were −21, −31.4 and −30 FUN: deliberately mediocre, because the objective punished blow-ups rather than rewarding lucky segments.

The offline holdout then replayed the final 300 drops per risk with a fresh 1,000 FUN bank and an empty pattern history. Results: low +0.3 FUN, medium +20.8 FUN, high −84.8 FUN. Two of three holdout segments were positive — an encouraging but small sample, and the high-risk segment still lost heavily. This split is predefined and offline, not an external preregistration, and 300 drops per risk is far too few to prove anything; it simply guards against the most obvious curve-fitting.

Across the full 1,000-drop history per risk, zero searched configurations were positive at all three risks simultaneously — the full-history positive count is 0, and the best full-history cell is null. The positives that do exist (medium fixed +4.8, medium Paroli +4.0, A100 high +123.2, C25 medium +139.2) are single-risk, single-history outcomes on a board with a built-in edge near 1%. They are interesting historical facts, not a method. If you want to explore them yourself, the free demo, the how-to-play guide and the project home are the right starting points — for learning and mathematics, not as any authorised French cash option.

• Selection used only the first 700 drops per risk; holdout used the final 300.
• Holdout: +0.3 low, +20.8 medium, −84.8 high.
• No configuration was positive across all three risks over the full history.