Martingale Mutiny

How it works

Your boat starts with 100 cargo and sails up to 30 legs toward an island at 250 cargo, with a shipwreck line at 15. Each leg one of two things happens. With probability 1 - e^(-λ) (λ = 0.1, about 10% per leg) a storm hits and your cargo is halved. Otherwise it is a calm-water leg: cargo moves by a fixed drift of +8 plus a random term σₙ·z, where z is a standard Normal draw and the spread grows every leg as σₙ = 4·(1.04)ⁿ - later legs are wilder than early ones. Cargo is floored at zero on a calm leg.

You have three moves each leg: sail on, drop anchor, or (for the first 3 legs only) ask to be walked through it. The guided option rolls the real outcome, shows you the exact formula with the actual σₙ and z draw substituted in, and asks you to compute the resulting cargo yourself before it is committed - graded to within 0.5 cargo units against the same rounded numbers you were shown.

The run ends three ways: cargo at or below 15 is a shipwreck, cargo at or above 250 reaches the island, and 30 legs completed means the voyage stops. Dropping anchor is the deliberate ending - but before your haul is tallied you must first predict the theoretical expected cargo E[Xₙ] at your current leg count, with no live readout to lean on. Only then do you see how your prediction, your actual haul, and the recursion compare.

How scoring works

The prediction at anchor is scored by relative error against the exact recursion value: within 5% earns 3 points (Excellent), within 15% earns 2 (Good), within 30% earns 1 (In the ballpark), and worse earns 0 (Off).

Progress tracking records an attempt under the optional-stopping skill: the anchor prediction counts as correct at 2 or more points, and each guided fill-in-the-blank counts as correct within 0.5 cargo of the formula's answer.

Reaching the island or shipwrecking is the random outcome of the voyage and never touches the graded score - only the two computations are graded. The result screen also compares your banked cargo to the theoretical E[Xₙ] and banners you as ahead of or behind theory.

The one-step recursion: conditioning on what can happen

The whole game reduces to one conditional expectation. Given cargo X with shock probability p per leg: with probability p the hold is halved (0.5·X), and with probability 1-p you gain μ = 8 on average (the σₙ·z term has mean zero, so it drops out of the expectation). So E[Xₙ | Xₙ₋₁] = (1-p)(Xₙ₋₁ + μ) + p·(0.5·Xₙ₋₁) = Xₙ₋₁·(1 - p/2) + (1-p)μ.

That is a linear recursion E[Xₙ] = a·E[Xₙ₋₁] + b with a = 1 - p/2 ≈ 0.9524 and b = (1-p)μ ≈ 7.239, starting from E[X₀] = 100. This move - condition on one step, take expectations, get a deterministic recursion - is the single most reusable technique in stochastic process interviews. Any Markov process with a small set of one-step outcomes yields to it.

Note what the recursion ignores: the growing σₙ and the boundaries. The Normal noise is mean-zero so it never enters E[Xₙ], and the recursion is for the unstopped process. The variance growth matters enormously for the distribution of outcomes - it decides how often you hit the island or the reef - just not for the expectation you are asked to predict.

Why the edge stops compounding: the equilibrium ceiling

Because the storm cost is proportional to your hold while the edge is a flat +8, expected cargo does not grow forever. The recursion has a fixed point x* = b/(1-a) = (1-p)μ/(p/2): set E[Xₙ] = E[Xₙ₋₁] and solve. With p ≈ 0.0952, that is about 152 cargo. Above the fixed point, the expected storm loss p·X/2 exceeds the expected calm gain (1-p)μ, so expectation is pulled back down; below it, the edge dominates and expectation rises.

This is the game's teachable core: proportional tail risk caps how much a flat additive edge can grow your expectation. Doubling your hold doubles what a storm takes but does nothing to what a calm leg gives. The island at 250 sits well above the ceiling of ~152, so reaching it means beating the odds, not the strategy - the expected path never gets there.

The convergence is geometric with ratio a ≈ 0.9524: each leg closes about 4.76% of the remaining gap to the ceiling. From 100, the gap is 52, so early legs add about 2.5 to expectation, shrinking every leg after that. This gives you a fast mental model - E[Xₙ] ≈ 152 - 52·(0.9524)ⁿ - that turns the anchor prediction into one exponentiation.

Stopping times and what dropping anchor really tests

Dropping anchor is a valid stopping time: the decision uses only what you have seen so far, never the next leg's outcome. The game makes you earn it by predicting E[Xₙ] blind first - which is exactly the discipline behind optional stopping arguments, where you reason about the expected value of a stopped process without peeking ahead.

The expectation logic says early legs are cheap to take (expectation rising fast toward the ceiling) and late legs are nearly free in expectation but expensive in variance: σₙ grows 4% per leg, so leg 25 has spread 4·(1.04)²⁵ ≈ 10.7 versus 4.16 on leg 1, and every leg carries the ~10% halving risk against the fixed shipwreck line at 15. Once your hold is large, a single storm costs you far more than the ~2 cargo the next leg adds to expectation.

A sound plan: sail while your hold is below the ~152 ceiling and comfortably above the reef, then treat further legs as paying variance for almost no expectation. When you do anchor, compute the prediction from the recursion rather than from your own cargo - your realized path is one noisy sample and frequently sits far from E[Xₙ], especially after a storm. The scoring is entirely about the recursion, not about your haul.

A worked example

You drop anchor after 5 legs. What is E[X₅]?

Step 1 - the coefficients. p = 1 - e^(-0.1) ≈ 0.0952, so a = 1 - p/2 ≈ 0.9524 and b = (1-p)·8 ≈ 7.239. The fixed point is b/(1-a) ≈ 152.

Step 2 - iterate from E[X₀] = 100. E[X₁] = 0.9524·100 + 7.239 ≈ 102.5. E[X₂] ≈ 0.9524·102.5 + 7.239 ≈ 104.8. E[X₃] ≈ 107.1. E[X₄] ≈ 109.2. E[X₅] ≈ 111.2.

Shortcut - gap form. E[Xₙ] = 152 - 52·(0.9524)ⁿ. At n = 5, (0.9524)⁵ ≈ 0.784, so E[X₅] ≈ 152 - 40.8 ≈ 111.2. Same answer, one calculation.

Step 3 - grading. Within 5% of 111.2 (about 105.6 to 116.8) earns 3 points; within 15% (about 94.5 to 127.9) earns 2 and still counts as correct for progress. Note the answer does not depend on whether you were stormed - your actual cargo after a storm might be 55, but E[X₅] is a property of the process, not of your path.

Common mistakes

• Predicting your own current cargo instead of the recursion's E[Xₙ]. The question asks for the theoretical expectation at your leg count; your realized path is one noisy draw around it.

• Compounding the edge forever - estimating E[Xₙ] ≈ 100 + 8n. That ignores the storms entirely and overshoots badly for large n; the true expectation never exceeds about 152.

• Applying the calm-leg drift of +8 as if it happened every leg. The unconditional per-leg gain is (1-p)·8 minus the expected storm loss p·X/2, which shrinks to zero as X approaches the ceiling.

• Letting the σₙ·z term into the expectation. z has mean zero, so the noise term vanishes from E[Xₙ] - it belongs in your risk assessment, not your prediction.

• Sailing on above the ceiling because the edge is positive. Above ~152 the expected storm cost exceeds the flat edge, so each extra leg lowers your expected cargo while the growing σₙ and the halving risk raise your chance of shipwreck.

Why interviews test this

One-step conditioning into a linear recursion E[Xₙ] = a·E[Xₙ₋₁] + b, then solving for the fixed point b/(1-a), is a staple of quant interviews - it is the same machinery behind gambler's ruin expectations, expected hitting times, and AR(1) processes, where the fixed point is the stationary mean.

The multiplicative-loss-versus-additive-gain structure is also a favorite: it is why interviewers ask about geometric versus arithmetic growth, Kelly-style sizing, and why a strategy with positive expected value per step can still be capped or ruinous. Being able to say 'the shock is proportional, the edge is flat, so expectation converges to (1-p)μ/(p/2)' - and to defend dropping anchor as a stopping time that uses no future information - covers both the computation and the concept an interviewer is probing.