Why every edge eventually gets arbitraged into a rounding error — shown on 2,669 trading days, not vibes.
The arc is always the same. You buy a $500 prop account, catch a run, pull out a few thousand. You blow a couple more accounts and you're still net positive, because one payout dwarfs the fees. It feels like skill.
Then the firm tightens the rules or slow-walks your withdrawal, and you move to a real broker. That's where the math catches up: no oversized payout to absorb the damage anymore — every losing trade is your own money, with nothing on the other side to win it back.
This isn't about discipline or psychology or a better strategy. It's colder: the market is structured so the average participant cannot win — not "usually doesn't," cannot, for the same reason a casino doesn't need to cheat. Any real edge attracts the traders and algorithms who also found it, and they compete it down to break-even, then to cost.
So I went looking in the data: ten years of DAX, 2,669 trading days, one-minute bars. The most basic question a trader can ask — can you tell, in real time, what kind of day you're in, and trade it for a profit?
Spoiler: no. And how it's no is more interesting than the no.
DAX index, one-minute bars from Dukascopy, 2015 → 2025, 2,669 trading days, regular cash session (09:00–17:30 Berlin). Everything below is computed from that — every chart is reproducible from the scripts at the bottom.
One honest caveat: this is a CFD on the index, and its "volume" is tick volume — a proxy, not real exchange volume. It changes nothing here (the questions are about price), but you should know it.
Almost every retail trading framework rests on one assumption: that a session has a kind. A trend day. A range day. A reversal day. The taxonomy itself is old and respectable — in Steidlmayer's Market Profile, Jim Dalton names these day-types as a descriptive vocabulary, a way to label a session after it has formed. That's a fair use, and not what I'm arguing with. The problem is the next move retail makes with it — recognize the kind early, switch to the matching playbook, win — which smuggles in two things the descriptive version never claimed: that the types are discrete to begin with, and that you can know which one you're in while it still matters. This section tests the first; Section 3 tests the second.
Take the first one. Before you can trade "trend days," they have to exist as a distinct thing. So I measured it. For each day I took the share of its daily range captured by the net move (|net move| ÷ daily range — near 1 when the day went one way cleanly, near 0 when it churned back to where it started), and plotted all 2,669.
It's one smooth hill. A Hartigan dip test (which checks for exactly the kind of valley you'd see between two real groups) returns p = 0.99: statistically, there is no valley. No gaps, no natural seams, no clusters. "Trend days" and "range days" aren't categories the market hands you — they're slices you draw, by hand, through a continuous gradient.
This is also why clustering algorithms "find" day-types and mislead people. k-means is mathematically flawless — that's the problem. Feed it a smooth black-to-white gradient and ask for two clusters: it splits down the middle, filing 49%-grey under "black" and 51%-grey under "white" — pixels your eye can't tell apart — while that 51% pixel sits in the same cluster as pure white. The split is valid and meaningless. k-means always returns k clusters; the boundary is the algorithm's, not the data's.
The market is a fluid. Everything below follows from accepting that.
If it's a continuum, the honest move is to score trendiness, not slap a label on it. And the gradient holds from every angle: sort the days from calmest to most one-directional, and three independent measures of trendiness climb together smoothly, with no step anywhere.
So we can absolutely measure how trending a day was. The question that decides whether you can make money is sharper: can you measure it while it still matters — early enough to act?
Here's the part the title promised.
I gave a model everything available by noon — the whole morning's price action, range, position, how choppy it had been — and asked it to predict the rest of the day (a clean split: the model only sees the past, the label only covers the future, so there's no cheating). Two questions, the obvious ones a trader needs answered: will the rest of the day trend, or just rotate?
Both answers land at a prediction skill of 0.53, where 0.50 is a coin flip. The character of the afternoon is, to a first approximation, unknowable from the morning.
And it gets worse — look at the right panel. When the model is most confident a day will rotate, it's right 36% of the time, below the 38% base rate. The predictions it's proudest of are the ones you should trust least.
This is the whole game. Every strategy that says "in a trending market do X, in a ranging market do Y" quietly assumes you can tell which one you're in. You can't. Not reliably, not in time. It's a coin flip with extra steps.
"But trend-following works," you say. "And mean-reversion works." Here's the uncomfortable truth: they both do — in exactly the place you can't reach.
Split the days by type after the fact and trade each playbook. Trend-following genuinely earns +0.59 R (+58 DAX points) on fast-trend days; fading genuinely earns +0.10 R on rotational days. The money is right there in the data.
They're perfect mirror images — each one profitable precisely where the other bleeds. And both are unreachable for the same reason as Section 3: you can't tell which bucket you're in until the day is over (that 0.53 coin flip again).
That label "after the fact" is doing heavy lifting, and it's where most backtests quietly lie to their authors. A day gets called a fast-trend day because it trended all day — including the part after you'd have entered. "Just trade the trend days" really means "just trade the days that worked," which requires a time machine.
And then there's the metric retail loves most: win rate. Watch what happens when you fade extremes and vary only the profit target:
Before costs, every version of the trade is break-even. After a single point of cost, every version loses. The punchline sits on the left: the tightest scalp wins 68% of its trades and still loses money. That equity curve looks gorgeous in a screenshot and bleeds in a brokerage statement.
Win rate is what you show investors. Expectancy is what the broker counts.
So why does this keep happening — to every retail trader, on every strategy, eventually?
Picture the skill of every market participant as a bell curve. Every trader and every algorithm is, collectively, an arbitrage machine: each one that finds a sliver of edge trades it away, dragging the whole distribution toward the average. The follow edge, the fade edge — both got ground down to ≈ 0 in the data because that machine already ate them.
What survives is a layer so thin that only the far tail can extract it — the players with the lowest costs, the fastest execution, or genuine private information. The consistently profitable aren't a little above average. They're way out in the tail, because the middle has been smoothed flat. (A calibration note, because honesty is the point: "3 sigma" is an illustration, not a measured constant — studies of day traders put the consistently profitable at roughly 1–3%, about 2–2.5σ.)
And the tail can never be smoothed all the way to zero — that's the Grossman–Stiglitz paradox. If markets were perfectly efficient, no one would be paid to gather information, so no one would, so prices couldn't stay efficient. A thin, irreducible edge has to remain to pay the few who are best at extracting it. Efficiency doesn't mean the edge is zero; it means it's exactly as small as it can be while still paying its best harvesters — and you are almost certainly not being paid.
This is also the answer to the prop-firm arc at the top. On a prop account the payout structure masks the math: one big withdrawal hides a string of blown accounts, and it feels like a system. A real broker strips the mask off. There's no asymmetric payout to hide behind — just your edge against everyone else's, and for the average participant that edge is a rounding error that costs turn negative.
Trading is a probability game barely distinguishable from a coin flip. To come out ahead over distance you don't need confidence, or a guru, or a cleaner chart — you need a structural edge over the other participants: lower costs, better execution, faster information, or a genuine statistical advantage that survives them all finding it too. Almost no retail setup has any of these.
And the cruelest part is built in: the moment a real loophole appears, it starts attracting the very people who close it. By the time it's on YouTube, it's break-even. By the time it's in a $500 course, it's a cost. The appearance of an edge guarantees its disappearance.
None of this says markets are random noise — they have rich structure, as Sections 1–4 show. It says the structure is inseparable from noise at the moment you have to decide, and costs eat whatever leaks through. The patterns are real. The edge is in the tail. Most of us — including the version of me that started this — are standing in the middle, flipping a coin and calling it skill.
Read this as a lab notebook, not a peer-reviewed paper — an honest first pass, open about its limits:
- One instrument. DAX only; results may not transfer to other markets.
- Proxy volume. Dukascopy CFD tick volume, not real contract volume.
- Cost assumption. A flat 1-point round-trip; no slippage or latency model (both make the picture worse, not better).
- Validation. There's an in-sample / out-of-sample split, but no rolling walk-forward.
- One setup family. Two playbooks tested out of an infinite space — which also means multiple-testing risk: torture enough variants and one looks profitable by chance.
None of these change the qualitative finding; all of them should temper any precise number.
.
├── README.md
├── dax_trend_expectancy.py # session segmentation, trendiness score, follow expectancy
├── step1_balance_detectability.py # Section 3: can you predict the day in real time? (AUC)
├── step2_fade_backtest.py # Section 4: unconditional fade expectancy
├── step2b_fade_sweep.py # Section 4: fade target-geometry robustness sweep
└── figures/
Data via dukascopy-node:
npx dukascopy-node -i deuidxeur -from 2015-01-01 -to 2025-06-01 -t m1 -f csv -v --cacheAll scripts are pure Python (pandas / numpy / scikit-learn / diptest / matplotlib).
Left: the share of days you call "rotational" depends entirely on where you draw the cutoff (26% at 0.25 → 51% at 0.45) — proof, in one chart, that the buckets are arbitrary slices of a gradient. Right: the regime mix is nonetheless remarkably stable year to year, which is partly why "what happened recently" adds nothing to the noon prediction — at the yearly scale there's little to predict. (Stability across years doesn't rule out week-scale clustering, which this view averages out.)
If this was useful or interesting, a ⭐ helps others find it.