A Good Model Pays the Patient: Why Consistency Is the Whole Game

If a trading model has a genuine statistical edge, the math says consistency is what converts that edge into money. No single month proves anything, because short samples are dominated by noise. The reward comes from letting the law of large numbers run: over enough repetitions, realized results converge toward the strategy's true expectancy. The hard part is not finding an edge. The hard part is continuing to take the signal during the stretches when the edge looks broken.

Key takeaways

• A statistical edge means positive expectancy per decision, not a win every time; a model with a 60% monthly win rate still loses four months in ten.
• A strategy that wins 60% of its months produces three losing months in a row about 6.4% of the time, which at monthly frequency is roughly once every year and a half.
• Morningstar's "Mind the Gap" research found fund investors earned on the order of 1 to 2 percentage points per year less than the funds they held, because they bought after hot stretches and sold after cold ones.
• The law of large numbers only guarantees convergence over many repetitions, so judging a monthly strategy on a 12-month sample is judging coin-flip noise.
• Consistency pays only when the edge is real; applied to a broken model, the same discipline compounds losses, so validation has to come first.

What does it mean for a model to be statistically good?

A statistically good model is one whose average outcome per decision is positive, measured across many decisions. That is the entire definition. It says nothing about the next decision. A model with positive expectancy is a weighted coin, not a crystal ball: weighted in your favor, still a coin.

This framing matters because it separates two things people constantly blur. The quality of a decision process is its expectancy. The quality of a single outcome is luck plus expectancy, and over one month the luck term is far larger. A bad month from a good model and a good month from a bad model look identical on a statement.

The cleanest working example of this is a casino.

Why do casinos never change the rules after a losing night?

Because the rules are the edge, and the edge is per spin, not per night. On an American double-zero roulette wheel there are 38 pockets and a single-number bet pays 35 to 1. The house's expected take is 2/38 of every dollar wagered, which is 5.26% per spin. That is the whole business.

And yet casinos have losing nights. Someone hits three numbers in a row, a table runs hot, and for a few hours the house is down real money. What the casino does about this is nothing. It does not widen the wheel, fire the dealer, or switch to a different game. It knows that 5.26% per spin times millions of spins is an arithmetic certainty in a way that 5.26% times forty spins is not. The casino's only strategy is volume. It wins by showing up tomorrow with the same rules.

A trader running a positive-expectancy model is the house. A trader who reacts to one bad month by rewriting the system is the house tearing up its own wheel because a gambler got lucky on a Tuesday.

Why do losing streaks happen to winning strategies?

Because that is what the binomial math demands, not because anything broke. Take a strategy that wins 60% of its months, which is a strong monthly hit rate. The chance of any given month starting a run of three straight losers is 0.4 × 0.4 × 0.4, about 6.4%. At twelve decisions a year, that means a three-month losing streak arrives roughly once every year and a half, on schedule, as a property of the win rate itself.

Stretch the horizon and the streaks stretch with it. Four straight losing months has probability 0.4^4, about 2.6%, so an 8-year record of 96 monthly decisions can be expected to contain a couple of them. Even a five-month cold streak, at roughly 1%, has decent odds of showing up once somewhere in those 96 months. Drop the win rate to a still-profitable 55% and all of these get more frequent. A four-month drawdown inside a multi-year track record is not evidence against the model. It is what the model's own statistics predicted in advance.

Quant GT's live track record is a concrete case: across 8 years of monthly five-stock portfolios, the model averaged roughly 58% per year, and that average emerged from a record that included losing months and multi-month cold stretches. The historical return was the sum of the whole sequence, rough patches included. Anyone sampling only the worst three months would have concluded the model was broken. Why the average landed where it did is covered in why Quant GT's returns have been so high.

How long before an edge shows up in results?

Longer than your patience wants, shorter than forever. The law of large numbers is the formal version: as the number of independent trials grows, the realized average converges to the true expectancy. The catch is the word "grows." Twelve monthly results are twelve coin flips. Flip a 60% coin twelve times and getting five or fewer heads, a losing-looking year from a winning process, is not rare. A single down year cannot distinguish a good model from a mediocre one.

At 36 months the picture sharpens. At 96 months, the length of an 8-year record, the noise has been mostly averaged out and what remains is close to the real expectancy. This is why serious quant shops evaluate strategies in years and why a monthly-rebalance system has to be judged on dozens of rebalances, not a handful. Short samples are noise. Long samples are the answer.

Why do most people quit at exactly the wrong time?

Because drawdowns feel like information and arrive when the urge to act is strongest. The standard failure pattern is strategy-hopping: abandon a system after a losing stretch, move to whatever performed best recently, then repeat when the new system hits its own inevitable cold streak. Run that loop and you systematically own each strategy's worst stretch and miss each one's recovery. Every paper drawdown becomes a realized loss.

This is not hypothetical. Morningstar's "Mind the Gap" studies measured the difference between fund returns and the returns fund investors actually earned, and found investors lagged their own funds by on the order of 1 to 2 percentage points per year. Same funds, same holdings. The gap came entirely from timing: money flowed in after strong performance and out after weak performance, which is buying high and selling low executed with great conviction.

A fixed systematic process exists to delete that option. A model that rebalances on a set monthly schedule, the way Quant GT's does, has no mechanism for quitting at the bottom, because no month's decision depends on how the previous month felt. The schedule is not a convenience. It is the behavioral firewall. Position sizing and exposure rules do the same job on the risk side, which is its own topic, covered in what is risk control.

When is consistency the wrong move?

When the edge was never real. Everything above is conditional, and the condition does the work: consistency converts a genuine edge into money, and it converts a fake edge into a steady drip of losses executed with perfect discipline. Patience is not a virtue in itself here. Patience applied to a broken model is just slow capitulation.

So the real question is never "should I stay consistent?" It is "have I verified the edge?" A backtest alone does not answer it. Backtests can be overfit, curve-matched to one regime, or quietly contaminated by survivorship bias, and an overfit backtest looks exactly like a real edge right up until live money arrives. What separates a validated edge from a mirage: out-of-sample testing, results that hold across periods the model never saw, and a live track record long enough for the law of large numbers to mean something. The statistical machinery for making that call rigorously is the subject of the hypothesis testing lesson, and it is the homework that has to be done before consistency earns its keep.

Get the order right and the rest is mechanical. Verify the edge first, with evidence that would survive a skeptic. Then hold the line through the losing months the math already promised you, because at that point the cold streak is not a warning sign. It is the toll the strategy charges on the way to its average, and the people who got paid the historical average were, definitionally, the ones still there when it arrived.

FAQ

What is a statistical edge in trading?

A statistical edge is positive expectancy: across many trades, the average outcome is a gain, even though plenty of individual trades lose. It is a property of the long run, not of any single result, which is why one month proves nothing in either direction.

How often do losing streaks happen to a winning strategy?

Constantly. A strategy that wins 60% of its months still produces three losing months in a row about 6.4% of the time, which at monthly frequency works out to roughly once every year and a half. Four- and five-month cold streaks fit comfortably inside an 8-year record.

How long does it take for a real edge to show up in results?

Longer than most people give it. A 12-month sample of monthly results is mostly noise; the law of large numbers only forces results toward true expectancy over many repetitions, so an edge becomes visible over several years, not several months.

When should you abandon a trading strategy?

When the evidence says the edge was never real, not when the equity curve feels bad. A drawdown that sits within the strategy's historical range is expected behavior; a model that fails out-of-sample testing or whose live results diverge sharply from its validated record is a different matter, and quitting it is correct.