Research
Where Risk Belongs
A stop-loss and a profit target are the two instincts everyone reaches for first. Tested against the one strategy in this project that carries a real edge, both destroy it — and for the same reason.
Both barriers cut the same tail.
Returns are right-skewed — a few large winners carry the entire average. A stop truncates them by ejecting you during ordinary noise. A target truncates them by design. Every loser survives both. The measured optimum for each is not to use it, and risk moves to position size instead.
A 2% stop turns the edge negative
The base strategy: enter at the opening auction after pre-market news, exit at the close. Adding a −2% stop and a +5% target, measured on 5-minute bars so the true order of events inside each day is known:
| outcome | frequency |
|---|---|
| hit the +5% target | 5.4% |
| stopped out at −2% | 34.7% |
| neither — exit at the close | 59.9% |
| configuration | mean | median | win rate | worst |
|---|---|---|---|---|
| with −2% / +5% barriers | −1.9b | −36.5b | 44.1% | −2.0% |
| no barriers, open to close | +36.6b | +22.4b | 55.6% | −8.6% |
The stop does not reduce the edge. It removes it, and then some. Note the shape of the trade-off: the barrier version has a better worst case — −2.0% against −8.6% — and a worse everything else.
2% is inside an ordinary day
| symbol | average daily range | a 2% stop is… |
|---|---|---|
| COIN | 5.51% | inside one ordinary day |
| PLTR | 4.66% | inside one ordinary day |
| TSLA | 4.26% | inside one ordinary day |
| NVDA | 3.40% | inside one ordinary day |
| AAPL | 2.11% | inside one ordinary day |
Every name in the universe, including the calmest. The stop is not catching disasters; it is catching Tuesday. You are removed from a third of positions by fluctuation that would have reversed, and you pay the full 2% each time.
Meanwhile +5% demands a move larger than a typical entire day’s travel, which is why it arrives 5.4% of the time. A 5.4% chance of +5% against a 34.7% chance of −2% is −0.42% expected from the barriers alone, before the edge gets a chance to work.
Every profit target makes it worse
Sweeping the take-profit level across a two and three day hold. Market-adjusted, net of 20 bps, with the sample split so the choice can be validated on data it never saw.
| take-profit | 2-day hold | 3-day hold |
|---|---|---|
| 2% | −5.8b | −8.5b |
| 3% | +6.1b | +1.1b |
| 5% | +19.5b | +13.6b |
| 8% | +30.3b | +24.5b |
| 10% | +34.6b | +29.2b |
| 1× daily range | +9.0b | +4.9b |
| 2× daily range | +27.6b | +21.8b |
| 3× daily range | +34.2b | +29.1b |
| none — hold to horizon | +41.7b | +41.4b |
The curve is monotonic. There is no interior peak to fit to, which is why this survives the usual objection: selecting on 2021–2024 picks “none” at both horizons and delivers +42.4 and +40.4 bps on 2024–2026. Nothing was optimised, because there was nothing to optimise.
One secondary result worth keeping: at comparable tightness, a target scaled to the stock’s own range beats a fixed percentage — 2× range returns +27.6b against +19.5b for a flat 5%. A fixed percentage means something different on COIN than on AAPL. Scaling makes it mean the same thing on both.
Both barriers cut the same tail
Measured skew of these returns is +0.65. The average is carried by a small number of large winners, and both barriers remove them.
A stop ejects you during ordinary fluctuation, before a winner has finished. A target caps a winner at the moment it is working. Neither touches the losers, which arrive intact either way. That is why the measured optimum for both is zero, and why the instinct producing them — small controlled loss, defined gain — is precisely inverted.
The barrier version genuinely does have a better worst case. It buys that with the entire edge.
Risk belongs in position size
Abandoning stops and targets is not abandoning risk control. It moves risk to the one variable that can actually be forecast.
Ranking these stocks by three-day return does not survive out of sample
(rho = +0.04). Ranking them by daily range does, emphatically
(rho = +0.83). You cannot predict direction. You can predict distance
— and distance is exactly what position size needs.
A third the size in COIN as in AAPL puts the same dollars at risk in each. That calculation is reliable in a way no stop level is, because it rests on the one quantity these markets keep repeating.
A $35,000 account taking one signal a day at full size, across 1,202 trading days: median $41,252 — but a 5th percentile of $16,082, and 37.8% of simulated histories ending below where they started, on a strategy with genuinely positive expectancy.
That is not the edge failing. It is one position at a time drowning a 22.8 basis point average in variance. Six positions a day at a sixth the size is the same edge with a fraction of the ruin risk. Breadth is the risk control that works here. The barriers only felt like one.
What would undercut this
Intraday ordering covers six months
The 5-minute test establishing the stop result runs Feb–Aug 2026. Daily bars extend the sample but cannot tell whether the high or the low came first, so the precise measurement is also the shorter one.
Thirty large caps in a rising market
Right skew, and the benefit of simply holding, may both behave differently through a sustained decline. This window does not contain one.
Costs are assumed, not observed
Twenty basis points per trade is an estimate. Every conclusion here is a comparison within the same assumption, so the ranking holds regardless — the absolute figures do not.
No barrier is not no risk
Dropping the stop moves the worst single trade from −2.0% to −8.6%. That is a real cost, paid in exchange for a larger average. An account that cannot survive the former should not be sized for the latter.
Entry at the opening auction following pre-market news, direction from headline classification. Barrier outcomes on 5-minute bars in true sequence; the take-profit sweep on daily bars, market-adjusted against SPY over the window actually held, with ties resolved against the strategy. Train Jun 2021 – May 2024, test Jun 2024 – Aug 2026.
This is a measurement of historical data. It is not investment advice, a forecast, or a recommendation about any security.