What the backtesting foundEMA200, 2000–2026
The board’s rule, replayed over today’s 220-name universe on daily bars and scored against random entries on the same tape. R is one unit of risk: the distance from entry to a 2 ATR stop.
- The daily EMA200 touch times entries better than chance. Held out of sample, none of 200 random books beat it.
- It still trails buy-and-hold: +108% against +160% over the same years, because it sits in cash while the universe compounds. The evidence is for alerting, not allocating.
- Long only. The short side loses money at every scale tested.
The timing edge, held out
Configuration picked on 2016–2021 by a declared rule, then scored once on 2022–2026. The control moves each entry to a random day within ±45 sessions on the same ticker; tightened to ±10 sessions the gap widens, to z = +4.62.
| Metric | Real | Random | z | Random beat it |
|---|---|---|---|---|
| Expectancy | +0.591R | +0.220R | +4.21 | 0 / 200 |
| Sharpe | 1.29 | 0.50 | +3.98 | 0 / 200 |
| Total return | +108.4% | +35.4% | +3.98 | 0 / 200 |
| Max drawdown | −23.7% | −26.0% | +0.69 | 49 / 200 |
1D against 1W
Same rule, 2007–2026, long only, each against random entries within ±10 bars. The weekly touch points the same way at about two standard deviations, fires 233 times in twenty years, and holds a book that sits two-thirds in cash: context for a daily signal, not a book of its own.
| Timeframe | Trades | Real | Random | z | Random beat it | CAGR | Exposure |
|---|---|---|---|---|---|---|---|
| 1D | 986 | +0.520R | +0.242R | 6.36 | 0 / 100 | 11.4% | 65% |
| 1W | 233 | +0.710R | +0.548R | 1.82 | 4 / 100 | 3.8% | 33% |
The short side
Negative at every span, and this comparison needs no control: it is two directions over the same bars.
| Span | Short trades | Expectancy |
|---|---|---|
| 8.8 years, 2017–2026 | 526 | −0.185R |
| 26 years, 2000–2026 | 1,231 | −0.353R |
| 63 years, 1963–2026 | 2,812 | −0.120R |
In a downturn
Each crisis split at the market’s own peak and trough. The short book is positive in every fall and negative in every rebound; over 26 years it gives back 91% of the account. Insurance whose premium costs more than the cover. The long book de-risks by accident: signals dry up in a fast crash, so it lost less than holding in the GFC and Covid.
| Fall | Long only | Short only | Buy & hold |
|---|---|---|---|
| Dot-comMar 00 – Oct 02 | −16.1% | +13.9% | −8.4% |
| GFCOct 07 – Mar 09 | −34.5% | +41.3% | −49.3% |
| CovidFeb – Mar 20 | −20.3% | +1.0% | −33.2% |
| 2022Jan – Oct 22 | −19.3% | +15.7% | −17.0% |
Read it with
- Survivorship, and it is severe: today's universe replayed backwards. The bias points up and grows the further back the replay reaches.
- One split date, one selected configuration. The random-entry control cancels most of the survivorship out of the edge figures; it cannot cancel it out of the return figures.
- 0 of 200 is the floor a 200-seed test can print, not a measured tail.
- Whether the rule picks the right names, and not only the right days, was never tested.
- No FX, no borrow cost, one look per bar, and no analyst overlay, which Yahoo does not publish a history of.