Independent quantitative research
Margin & Co. tests the four indicators retail traders rely on most — EMA, VWAP, MACD and RSI — on data they have never seen, after spreads, commission and slippage. The results are published weekly, including the weeks nothing works.
Week 03 · 27 August 2026
None of the 4 indicators beat buy & hold once costs were applied.
Tested on 756 S&P 500 and FTSE 350 companies, measured over 8 August 2023 to 27 August 2026 — data the rules had never seen. Buy & hold returned a Sharpe ratio of 1.53 over the same window. MACD, VWAP, EMA produced a negative Sharpe after costs — they lost money net of what it cost to trade them.
The decay curve
Results
756 companies — MMM, AOS, ABT, ABBV, ACN, ADBE and 750 others — over 17 August 2016 to 27 August 2026, costed at 8 basis points per side.
| Indicator | In-sample Sharpe |
Out-of-sample Sharpe (gross) |
Out-of-sample Sharpe (net) |
Trades | Cost paid | Max drawdown | Verdict |
|---|---|---|---|---|---|---|---|
| RSI | 0.12 | 0.55 | 0.48 | 85982 | 9.4% | -6.4% | WEAKENED |
| MACD | 0.15 | -0.36 | -0.63 | 247525 | 52.3% | -25.6% | FAILED |
| VWAP | -0.19 | -0.30 | -0.69 | 375820 | 79.4% | -27.5% | FAILED |
| EMA | -0.17 | -0.77 | -0.84 | 59951 | 12.6% | -30.8% | FAILED |
| Buy & hold | 0.85 | 1.53 | 1.53 | 0 | 0.1% | -14.3% | benchmark |
Why this exists
Search for “MACD strategy backtest” and you will find thousands of results showing spectacular returns. Almost all of them share the same three flaws.
Try enough combinations of moving-average lengths and one of them will have worked, by chance. That is a search result, not a discovery. Here the rules are fixed on the first 70% of history and the remainder is measured once.
Every trade pays the bid–ask spread, commission and slippage. A strategy that trades daily pays those costs roughly 250 times a year. In week 03, the worst offender paid away 79% of its notional value.
A signal computed from Monday’s closing price cannot be traded until Monday has closed — but most homemade backtests let it earn Monday’s return anyway. That single mistake is worth about 20 points of Sharpe ratio in our own test harness, which is why a test exists specifically to catch it.
Survival Score
3,308 indicator-and-stock combinations were tested this week. Each was scored from 1 to 10 on whether it still made money on data it had never seen, after paying realistic trading costs. 10 scored 8 or above, meaning they survived convincingly. 1,294 of 3,308 finished with a positive Sharpe ratio after costs, though a positive Sharpe is a low bar and says nothing about whether the rule beat simply owning the stock. The median score was 2.4. A low median is the expected result rather than a broken test: most published indicator backtests do not subtract trading costs or check performance on unseen data, and doing both is what removes the apparent edge. Treat the 10 survivors with suspicion rather than excitement. Testing 3,308 combinations and reporting the best of them is precisely how a search produces results that do not repeat — at 0.3% of everything tested, this is close to what pure chance would throw up. A combination is only interesting if it survives again next week, on data that did not exist when it was selected.
Survived
In-sample Sharpe 1.57, out-of-sample net of costs 1.66, positive in 100% of walk-forward windows, worst drawdown -28.2% across 39 trades.
How the score works. Each indicator-and-stock combination is scored 1 to 10: 60% out-of-sample Sharpe ratio net of costs, 25% how consistently it held up across rolling test windows, 15% worst drawdown. A rule that lost money after costs is capped at 3, and one that barely traded is capped at 5 — otherwise sitting in cash would score well for doing nothing. The full formula is set out here, so you can disagree with a weight and recompute the table yourself.
The week
3,308 combinations were tested this week. 10 scored 8 or above. The median was 2.4, and 1,294 finished with a positive out-of-sample Sharpe ratio after costs.
| Indicator | Median score | Survived (8+) |
|---|---|---|
| EMA | 2.1 | 0 |
| MACD | 2.2 | 4 |
| RSI | 3.8 | 4 |
| VWAP | 1.7 | 2 |
Against last week. 10 of last week's 12 survivors survived again. Repeating twice is weak evidence, not strong: with this many combinations tested, some will repeat by chance alone. These two runs are only 3 days apart, so they were scored on almost the same data. A combination repeating across windows that overlap this heavily is close to arithmetic rather than evidence, and this comparison should not be read as one until the runs are a full week apart.
Note
NVIDIA reported second-quarter results on 26 August: revenue of $96.2bn, up 106% on a year earlier, with data centre revenue of $89.0bn, up 117%. It guided the current quarter to $108bn. By any measure the business had an extraordinary year.
Here is what the four indicators scored on NVDA in this week's run:
| Indicator | Score | OOS Sharpe after costs | Verdict |
|---|---|---|---|
| EMA | 3.8 | 0.18 | Did not survive |
| VWAP | 3.4 | 0.24 | Did not survive |
| MACD | 2.4 | -0.08 | Did not survive |
| RSI | 1.0 | -0.96 | Did not survive |
RSI scored 1.0, which is the floor. That is the same rule the live paper portfolio trades: buy when the fourteen-day RSI drops below 30, sell when it recovers through 50. Applied to NVDA, out of sample and after costs, it produced a Sharpe ratio of −0.96 — it lost money, reliably, across a period in which the company more than doubled its revenue.
That is the whole argument of this project in one stock. The indicator has no opinion about NVIDIA. It cannot see revenue, margin, or guidance. It sees a number derived from recent prices, and a rule that buys dips in a stock that mostly went up will sell the winners early and buy the falls that keep falling. Being right about the company and being right about the rule are different things, and only one of them is being tested here.
This is not a view on NVIDIA. Nothing here says the shares are cheap, expensive, or worth owning. It says that four specific technical rules, tested the same way as every other stock in the universe, did not survive costs on this one. A Survival Score is a statement about a rule's past behaviour, never a forecast about a share price.
Next week
Every one of these is re-scored next week on a window that includes days which did not exist when the score was assigned. No result is re-used and nothing is carried over.
| Indicator | Ticker | Score | OOS Sharpe |
|---|---|---|---|
| VWAP | MNTN.L | 9.2 | 1.66 |
| RSI | INVH | 9.2 | 1.63 |
| RSI | MUT.L | 8.8 | 1.32 |
| MACD | WIX.L | 8.7 | 1.18 |
| MACD | DATA.L | 8.4 | 1.70 |
| RSI | ABNB | 8.4 | 1.45 |
| MACD | SSIT.L | 8.4 | 1.29 |
| MACD | MNTN.L | 8.4 | 1.26 |
| RSI | UBER | 8.3 | 1.51 |
| VWAP | SSIT.L | 8.0 | 1.28 |
If these were real, most should still clear 8 next week. If they were noise, most should fall away. Either outcome is published, and the second is the more likely one: test thousands of combinations and a handful will look good by accident.
Not a forecast. This is a list of claims being checked, not a list of stocks to buy. Nothing here is a prediction about any company, and a high Survival Score is a statement about a rule's past behaviour after costs, not about what a share price will do next.
The weekly report
Which indicators survived, which didn't, and how the live paper portfolio did. One email a week. No advertising, and your address is never shared with anyone.
What you are agreeing to, precisely. Your email address, the date, and the fact that you consented are stored — nothing else. No name, no location, no tracking. The list is held on this site's own infrastructure and is not shared with any mailing company, advertiser, or anyone else. Unsubscribing deletes the record rather than flagging it inactive.
Archive
No indicator beat the benchmark. 763 companies, out-of-sample 2 August 2023 to 24 August 2026, costed at 8bps per side.
No indicator beat the benchmark. 763 companies, out-of-sample 1 August 2023 to 21 August 2026, costed at 8bps per side.