Strategy · Capital
The Heisenberg Principle of Trading
Strategy design, validation, and capital allocation in adversarial markets.
Strategy · Capital
Strategy design, validation, and capital allocation in adversarial markets.
In early August 2007, the most sophisticated market-neutral funds on earth started losing money at speeds their models called impossible. Goldman’s flagship quant funds reportedly fell double digits inside days; the biggest statistical-arbitrage books on Wall Street were hit in the same hours — while the broader market barely moved. Goldman’s CFO famously described “25-standard-deviation moves, several days in a row”: the models weren’t just losing; they were watching something they had assigned probability zero.
What happened was simpler, and stranger. Years of quiet success had led the smartest firms to nearly identical positions — the same signals, mined from the same data, crowded into the same trades. When one large book was forced to unwind, its selling moved prices against everyone else’s identical positions, forcing more unwinds, moving prices further. The strategies weren’t broken; the market they measured had become mostly made of each other. And within days most of those strategies snapped back — having ruined whoever de-levered at the bottom.
No model saw it coming, because the models were the cause. Measure a market, trade on the measurement, and you change the thing you measured; let everyone measure alike, and everyone changes it together. That isn’t bad luck, and it isn’t a bug in one firm’s backtest. It’s something much closer to a law of physics — and it has a name.
Trading strategies are cheap. Anyone with a data feed and a notebook can produce a promising one by Friday. What’s expensive is evidence — finding out whether an edge survives contact with a live market that reacts to being traded — and what’s genuinely scarce is the discipline to spend capital on that question in the right order. This piece is about that discipline: where strategies come from, how they earn capital, and how they lose it.
Start with the physics. Heisenberg’s uncertainty principle says you cannot pin down both a particle’s position and its momentum with arbitrary precision — the sharper one gets, the blurrier the other. In the popular telling, the act of measurement disturbs the thing being measured.
Markets run on the same constraint: you cannot simultaneously know the true price of an asset and execute meaningful size at that price. Want price precision? Trade small. Want your full size done now? Cross the spread and accept that the price you got is not the price you saw. The act of trading moves the market you’re trying to measure.
Physicists will object that the uncertainty principle is about complementarity — pairs of properties that cannot jointly be sharp — not about clumsy instruments, and they’re right. The market version holds on both counts. Price and size are a genuine conjugate pair: precision in one costs precision in the other, no matter how good your execution is. And the measurement genuinely disturbs: every order you place discloses intent to counterparties who are paid to read it. Your order flow is information — theirs.
Three things follow, and everything else in this piece is built on them:
You can watch the constraint operate in the order book itself. Displayed liquidity is illusory — depth that looks available disappears when you reach for it, because reaching for it is information too. Execution is a standing trade-off between going fast (impact) and going slow (drift). There is no setting where the uncertainty goes away.
Which leads to the core insight, and it is exactly the reason promising backtests die: signal research assumes you are a passive observer. Live execution makes you an active participant. The edge you measure in simulation is not the edge you capture in production — not because your backtest was sloppy, but because the measurement itself changes the answer.
Most systems are designed to fight this — more precision, more parameters, more confidence. They lose. The design mantra we run on instead: bounded confidence, controlled ignorance. Trying to eliminate uncertainty accelerates edge decay. The job is to manage it, not collapse it.
If uncertainty can’t be collapsed, it has to be priced. That’s what a strategy lifecycle is for.
Design → validate → allocate → monitor → kill. Each stage is a filter that most candidates should not survive, and each stage earns the next increment of capital. The rule underneath is simple: capital at risk grows only as uncertainty shrinks — and because of the measurement problem, uncertainty never reaches zero, so allocation is never finished. It’s a continuous decision, not a graduation ceremony.
Most firms run the inversion. A strategy arrives with a champion, conviction substitutes for evidence, and it launches at full size — maximum capital at maximum uncertainty. Then reality arrives on its own schedule, and the strategy is killed not by data but by pain: a drawdown large enough to force the decision that evidence should have made months earlier, at a fraction of the cost.
Fund on conviction, kill on pain — that’s the default. The discipline is the opposite: fund on evidence, kill on data.
Everyone in this business talks about alpha — finding an edge, protecting an edge, mourning a decayed one. Almost nobody asks where edge comes from in the first place. It doesn’t come from mining patterns until one looks profitable — a pattern with no reason to exist has no reason to persist. It comes from market structure: for a strategy to earn money repeatedly, someone must repeatedly pay it. So the first question we ask of any candidate is not “does it backtest?” but “whose money is this, and why is it still on the table?”
That question has a complete answer set — a partition, not a brainstorm. A counterparty accepts a losing trade for exactly three reasons:
They don’t know. The counterparty is less informed than you, or slower to act on what’s knowable. You hold an information edge — directional prediction, stat-arb, anything that forecasts. How it dies: it decays as others learn what you know; the more visible and the more crowded, the faster.
They know, and pay anyway. You provide something they value — immediacy when they need to trade now, or taking risk off their book. You’re earning a risk premium — market making, carry, short-horizon reversion. How it dies: it mostly doesn’t. This is the one root that isn’t an inefficiency — it’s the price of risk, and it survives even in an efficient market, for whoever can bear the risk cheapest.
They’re forced. Mandates, liquidations, redemptions, hedging schedules — flow that must trade regardless of price. You’re supplying capacity to structural flow — arbitrage, rebalance capture. How it dies: competed down to the cost of the race; the discrepancy persists, but capturing it pays only the racing cost.
Every strategy anyone names is one of these three wearing instrument-and-horizon clothing. And because the roots die differently, the root a candidate draws from tells you — before any backtest — how it will decay and who you’re racing. If the answer to “whose money is this” is “someone slower will keep losing it,” go back to the measurement problem: that edge is a race, and races get faster.
Knowing the root is where design starts; surviving the decay it predicts is what the rest of design is for. Four principles:
A signal is a distribution, not a number. Point predictions and fixed thresholds are overconfidence encoded. A signal worth running carries its own uncertainty and its own half-life — an estimate of how fast it decays — and position size should shrink as either deteriorates. The corollary changes how death is detected: the kill trigger is the rate of divergence between live results and expectation, not P&L alone. P&L is noisy and slow; divergence is the earlier, cleaner witness. This is the strategy-level twin of the engine kill switch from Building a Trading Engine.
Underfit on purpose. In-sample performance is cheap to buy and expensive to own: every parameter added to fit history better is a bet that history repeats in that dimension. We penalize complexity explicitly and accept lower in-sample numbers. What that buys is out-of-sample survival — the only sample that pays.
Execution is information control, not cost minimization. If your behavior is deterministic, it’s a signature, and signatures get read. Within bounds, entropy is protective: vary timing, vary slice size, vary venue. The bound matters — randomize the execution, never the thesis. Sometimes this costs a little in slippage. It’s the premium on alpha longevity, and it’s usually cheap.
Build alphas as modular, replaceable units. If every edge decays — and the measurement problem says it must — then the system should be a portfolio of small, semi-independent alphas on a shared risk and execution layer, designed to rot gracefully and be swapped without surgery. The alternative is the god model: one monolith that’s brilliant until it’s wrong, and can’t tell you which part went.
Validation is a purchasing problem: each stage buys information about a strategy, at a price.
A backtest is the cheapest information available — and the least trustworthy, because it is exactly the passive-observer measurement the market punishes. Paper trading costs calendar time and buys contact with live market microstructure — real prices, real timing, simulated fills. Minimum-quantity live costs real money and buys the only evidence that settles the question at short horizons: real fills, real queues, real adverse selection. Fees and fills are where we’ve watched entire strategy classes die, so for high-turnover strategies, fills are the edge — everything before them is estimation.
The ordering rule: never buy expensive evidence when cheap evidence can kill the idea first. Run the funnel backtest → paper → min-qty live, and let each stage veto before the next spends more.
The failure mode that consumes entire teams starts right here: optimizing against an instrument nobody has validated. Months of engineering and compute go into training models and tuning strategies against a backtest — without ever establishing whether the backtest itself is right, or how close its fills, fees, and timing come to live behavior. Tune hard enough against a naive simulator and you will find “edge”; sometimes it’s a deep-learning model, expensively fitted to predict what is — net of costs — a random walk. Then comes the cruelest step in the sequence: the first live trades catch a favorable market and print money — beta, dressed as skill — and euphoria does the rest. The team concludes the model has struck gold, scales fast, and meets an ending that was predictable from the day the backtest went unexamined. The instrument you optimize against deserves more scrutiny than the thing you optimize: an uncalibrated measurement device doesn’t produce edge. It produces conviction.
Which points at the asymmetry, the most misunderstood thing in the sequence: a backtest can kill a strategy; it cannot validate one. Failing in simulation is disqualifying — the passive measurement flattered you and you still lost. Passing in simulation proves only that the flattering measurement was passed. Validation runs one direction: only live fills validate. A backtest earns a second, better life later — once calibrated against your own recorded execution reality, it stops being a gatekeeper and becomes an acceleration tool, a way to iterate fast between live validations. But calibration data comes from live trading. There is no shortcut around the market.
Run this funnel across many candidates and a bookkeeping problem appears: strategies at different stages, tested on different markets, with different sample depths — how do you compare them when capital is one pool? Our answer is a composite score, 0 to 100, built from four deliberately boring dimensions: beta-adjusted P&L (alpha dollars after stripping out market exposure), win rate, profit factor, and Sharpe on beta-adjusted returns. Nothing exotic — that’s the point. Each dimension normalizes against fixed bounds, so a 78 means the same thing this quarter and next, and the beta adjustment keeps a bull market from dressing exposure up as skill. The number isn’t the insight; the discipline of scoring every strategy the same way is.
Run that funnel for years and the most valuable output isn’t the strategies that passed. It’s the map of what reliably fails — conclusions bought with real capital and real months, which is the only currency that buys them.
Some of ours: short-horizon directional prediction on the major pairs — trend, breakout, reversal — did not survive our validation; the patterns are real in-sample and gone at the fill. Passive tight-spread market-making on majors fails structurally: adverse selection eats the spread you think you’re earning. And funding carry on majors is real but small — on the order of 3–4% net unlevered, not the 15–25% marketed; the difference is leverage, turnover, and scale doing the work the edge can’t.
None of that is a coincidence. Read against the three-root partition above: directional prediction and tight-spread market-making on majors sit on the information and structure roots — the two that decay to the cost floor first — deployed on the most efficient, most crowded instruments in crypto. The map isn’t a list of bad ideas; it’s the taxonomy doing exactly what it predicts.
The directional entry deserves nuance, because trend-following famously does work — somewhere. It’s one of the oldest documented premia in finance, and whether it pays depends on everything the label hides: horizon (the days-to-months trend premium is a different animal from intraday prediction), breadth (it’s harvested across dozens of markets, not two), regime (it earns in sustained moves and bleeds in chop), and costs (a marginal signal at institutional fees is a dead one at retail fees). What our validation killed was the specific corner we could actually trade: short-horizon, price-derived trend prediction on the major crypto pairs, at our costs and our scale. Version after version through the funnel, it never produced an edge our live fills could confirm. And the honest reading keeps some humility: perhaps the edge is there for someone with information we don’t have — flow visibility, positioning data, breadth we can’t reach. Both readings end in the same decision. A strategy that can’t prove its edge on your desk — with your costs, your data, your scale — is dead for you, and “dead for us, in this corner, at our costs” is exactly the resolution at which the map is drawn.
We publish conclusions, not methods — the map is the asset, and how it was drawn is the moat. But the conclusions alone are worth someone’s next four months: every entry above is a strategy class that much of the market is still actively paying to rediscover.
Crypto sharpens all of it. Thinner liquidity means the measurement problem bites at smaller size. Faster imitation means shorter half-lives. Reflexivity means regimes turn harder and historical calibration expires sooner. A system that chases precision dies everywhere — in crypto it just dies faster.
Sizing is where the philosophy either becomes real or becomes decoration.
The textbook answer is Kelly: size proportional to edge over odds. But Kelly assumes a stable edge and a known distribution, and we spent four hundred words above on why you have neither. So the default is anti-Kelly: fractional sizing against your own estimates, hard caps per strategy regardless of how good the numbers look, and allocation that shrinks automatically as crowding rises or the edge’s own volatility widens. Uncertainty means less leverage, not more modeling. The instinct to answer doubt with a more refined estimate is exactly the instinct that ends firms.
Capacity gets the same treatment. As Building a Trading Engine argued, capacity is endogenous — it shrinks as you scale, as others copy, as volatility shifts. So it’s watched, not assumed: track marginal impact as size grows, throttle before the market does it for you, and when a strategy wants more capital than its capacity supports, the answer is another alpha, not a bigger position.
And the kill discipline closes the loop. Monitoring against expectation — the divergence-rate trigger from the design principles — retires strategies on data, at small size, before P&L forces the decision at large size. Why Trading Firms Die called concentration in a decaying edge the slow death, death by years. This process is what prevents it: the portfolio sheds decaying edges the way it sheds losing trades — routinely, and without a meeting.
Step back and the pieces assemble into a shape.
Backing a single strategy — or a single-strategy manager — is a concentrated bet on a decaying asset, run by someone whose incentive is to believe in it past its expiration. Every element of this piece points the other way: modular alphas because each one dies; a validation funnel because most candidates should die cheaply; scored graduation because capital is one pool; anti-Kelly sizing because confidence is the enemy; kill discipline because the deaths must be routine.
What that adds up to isn’t a book of positions. It’s a factory: a process that originates candidate edges, validates them against real fills, allocates by evidence, and replaces them as they decay. The individual strategy stops being the asset. The pipeline is the asset — the strategies are inventory with a shelf life.
This is where the three pieces of this series converge. Survivability demands diversification (Why Trading Firms Die); the engine makes running many strategies operationally possible (Building a Trading Engine); and the lifecycle in this piece is what keeps the portfolio stocked with edges that have earned their capital. A durable trading operation isn’t one great strategy run well. It’s a system that doesn’t depend on any particular strategy staying alive.
A self-diagnostic, from our internal checklist. If your system:
— then it is fighting the measurement problem instead of managing it, and the market will eventually collect. The mantra we design against: optimize for edge longevity under adversarial observation, not predictive accuracy under static assumptions.
If you’re a quant with a strategy that has never met a real fill — that’s the exact gap the graduation funnel exists to close, and it’s what our validation platform does: backtest, paper, then min-qty live on our capital, with your IP staying yours. If you allocate capital, the diagnostic is one question: ask a manager what evidence gates their capital — and when they last killed a strategy on data alone. The answer tells you whether you’re backing a strategy or a system. That’s the conversation we’re set up to have.
Either way: what’s the most expensive lesson a backtest ever taught you live? We read every reply — and the death-proof map says most of those lessons are still being paid for in parallel, right now, across the market.
Martian Mobile is a proprietary crypto trading firm operating since 2018. This is part of a series on survivability, building trading operations, and validating edges — the full series lives here.