Verdict: Rejected · Failed at: Out-of-sample
Thin edge that does not survive the second half: on the Nasdaq 100 the default 9.1 made almost all of its money in 2014-2021 and lost in 2022-2026; 9.2 lost money on the S&P 500 in every setting tested (hypothetical backtest).
Hypothetical results from a backtest on broker CFD prices, after spread, commission and overnight financing. Not a forecast and not advice.
The numbers
| Metric | Nasdaq 100 | S&P 500 | Gold |
|---|---|---|---|
| Trades | 238 | 219 | 156 |
| Winning trades (%) | 35.3 | 33.8 | 29.5 |
| Profit factor | 1.20 | 1.23 | 1.39 |
| Total return (% of position, not compounded) | 39.9 | 32.8 | 38.5 |
| Maximum drawdown (%) | 25.3 | 19.9 | 30.0 |
| Return / drawdown | 1.58 | 1.65 | 1.28 |
| Return 2014–2021 (%) | 43.0 | 18.5 | -18.8 |
| Return 2022–2026 (%) | -3.1 | 14.3 | 57.3 |
| Time in the market (%) | 49.1 | 44.8 | 30.3 |
| Longest losing streak (trades) | 14 | 11 | 10 |
| Buy and hold, same period (%) | 716.5 | 298.7 | 235.0 |
| Buy and hold, maximum drawdown (%) | 35.5 | 35.4 | 26.6 |
Hypothetical backtest results, % of position, not compounded. Buy-and-hold columns are historical price changes.
Default settings: setup (1 = 9.1, 2 = 9.2): 9.1; EMA length (days): 9. 2014-06-01 → 2026-10-01.
Year by year (%)
| Year | Nasdaq 100 · strategy | S&P 500 · strategy | Gold · strategy | Nasdaq 100 · buy & hold | S&P 500 · buy & hold | Gold · buy & hold |
|---|---|---|---|---|---|---|
| 2014 | 6.5 | -4.8 | -1.7 | 13.4 | 7.0 | -2.0 |
| 2015 | -1.8 | -2.6 | -4.3 | 9.2 | 0.0 | -13.2 |
| 2016 | 1.8 | 1.5 | 9.6 | 5.0 | 8.5 | 8.6 |
| 2017 | 9.1 | -5.7 | -0.5 | 31.5 | 19.4 | 13.1 |
| 2018 | -4.6 | 3.6 | -2.5 | -0.8 | -6.0 | -1.5 |
| 2019 | 8.9 | 9.8 | -7.1 | 38.0 | 29.0 | 18.3 |
| 2020 | 21.3 | 14.1 | -8.8 | 47.3 | 16.1 | 25.1 |
| 2021 | 1.9 | 2.6 | -3.4 | 26.9 | 27.1 | -3.7 |
| 2022 | -9.4 | -11.7 | 0.6 | -32.9 | -19.5 | -0.3 |
| 2023 | 21.7 | 4.3 | 12.8 | 53.5 | 24.0 | 13.1 |
| 2024 | 3.0 | 12.9 | 16.1 | 25.0 | 23.5 | 27.2 |
| 2025 | -0.5 | 1.9 | 17.2 | 19.9 | 16.2 | 64.6 |
| 2026 (to September) | -17.9 | 6.9 | 10.6 | 21.0 | 12.1 | -3.3 |
Hypothetical backtest results, % of position, not compounded. Buy-and-hold columns are historical price changes.
Every setting we tried
Nasdaq 100
| setup (1 = 9.1, 2 = 9.2) \ EMA length (days) | 7 | 9 | 12 | 15 |
|---|---|---|---|---|
| 9.1 | 1.23 · 33.9 / 16.9 | 1.20 · 43.0 / -3.1 | 1.28 · 46.0 / 3.4 | 1.34 · 47.1 / 5.2 |
| 9.2 | 0.86 · 7.0 / -11.7 | 0.77 · 5.5 / -14.2 | 1.20 · 8.0 / 0.5 | 1.15 · 7.4 / -0.2 |
Hypothetical backtest. Cells: profit factor · return 2014–2021 / 2022–2026 (%). Settings that made money in both halves: 4 of 8.
S&P 500
| setup (1 = 9.1, 2 = 9.2) \ EMA length (days) | 7 | 9 | 12 | 15 |
|---|---|---|---|---|
| 9.1 | 1.09 · -0.7 / 14.7 | 1.23 · 18.5 / 14.3 | 1.14 · 21.2 / -1.4 | 1.16 · 4.3 / 16.4 |
| 9.2 | 0.62 · -12.9 / 0.8 | 0.91 · -3.8 / 0.9 | 0.85 · -5.8 / 0.6 | 0.81 · -6.6 / -1.1 |
Hypothetical backtest. Cells: profit factor · return 2014–2021 / 2022–2026 (%). Settings that made money in both halves: 2 of 8.
Gold
| setup (1 = 9.1, 2 = 9.2) \ EMA length (days) | 7 | 9 | 12 | 15 |
|---|---|---|---|---|
| 9.1 | 1.29 · -14.6 / 44.9 | 1.39 · -18.8 / 57.3 | 1.23 · -5.7 / 27.9 | 1.25 · 3.2 / 21.0 |
| 9.2 | 1.98 · 3.7 / 8.1 | 2.49 · 2.6 / 21.2 | 1.63 · -3.1 / 17.0 | 1.50 · -0.5 / 12.6 |
Hypothetical backtest. Cells: profit factor · return 2014–2021 / 2022–2026 (%). Settings that made money in both halves: 3 of 8.
The rules we tested
- 9.1: when the exponential moving average of daily closes turns up, place a buy stop one tick above that day's high; protective stop at that day's low.
- 9.2: while the average is rising, if a day closes below the previous day's low, place a buy stop above that day's high; same stop.
- Exit (our simplification): the close of the day on which the average turns down, or the protective stop. Many traders instead exit when that day’s low is broken.
Rules attributed to Larry Williams (as commonly taught in Brazil); summarised. This is not one of our strategies.
Cite this test
Empirical Quant, “Larry Williams 9.1 and 9.2 setups: backtest 2014–2026”, tested 2026-10-11, https://empiricalquant.com/backtests/larry-williams-9-1/. Data: https://empiricalquant.com/data/graveyard/larry-williams-9-1.json (CC BY 4.0).
What the 9.1 setup is
The 9.1 is a long-only trend entry attributed to Larry Williams, although the numbered setup names (9.1, 9.2 and so on) come from how the setups are taught in Brazil rather than from his own publications. It is very popular among retail traders in Brazil and watches a 9-day exponential moving average of daily closes. When the average turns from falling to rising, the trader places a buy stop just above that day's high, with a protective stop at that day's low, and exits after the average turns down again. The 9.2 is a pullback version, as we implemented it (definitions of the 9.2 vary between teachers): while the average is rising, a day that closes below the previous day's low sets up the same buy stop.
We tested the rules as they are commonly taught, on daily bars, without adding filters. One simplification: we exit at the close of the day the average turns down, while many traders instead exit when that day's low is broken. The point of this page is to answer one question with data: does the setup have an edge after real trading costs, and is that edge stable over time? It is not one of our strategies, and we do not trade it. Our own process is described on the method page, and this test is listed in our strategy graveyard.
Why we rejected it
Three problems with the default 9-day setting, each enough on its own.
First, the profit is small next to the risk. On the Nasdaq 100 the default 9.1 needed 238 trades over 12 years to make a hypothetical +39.9%, and on the way it fell 25.3% from its peak in the simulation. A hypothetical profit factor of 1.20 leaves very little room for slippage, wider spreads or a slightly different broker feed.
Second, the edge is not stable over time. In the backtest, on the Nasdaq 100, the default 9.1 made a hypothetical +43.0% from 2014 to 2021 and then lost 3.1% from 2022 to September 2026, a stretch in which the index itself rose strongly. On gold the pattern is reversed: −18.8% in the first half and +57.3% in the second, which mostly reflects gold's own rally after 2022. A rule that only works in one half of the sample is following the market, not finding an edge.
Third, it falls far behind simply holding the market. Over the same period the Nasdaq 100 price rose 716.5% and the S&P 500 298.7%. Even if you add up the Nasdaq 100's yearly returns without compounding, buy-and-hold gives about +257%, more than six times the 9.1 backtest result. The setup was in the market less than half of the time, so it missed a large part of the big up years.
What changing the settings does
A real edge should survive small changes to its settings. We ran every combination of setup (9.1 or 9.2) and moving-average length (7, 9, 12 and 15 days) on each market, and the full grid is in the tables above. On the Nasdaq 100, 4 of 8 settings made money in both halves of the simulated sample; on gold 3 of 8; on the S&P 500 only 2 of 8. Every 9.2 setting on the S&P 500 lost money overall.
The 15-day 9.1 was the only setting that made money in both halves on all three markets, but the margins were thin: +5.2% on the Nasdaq 100 in 2022–2026, +4.3% on the S&P 500 in 2014–2021 and +3.2% on gold in 2014–2021, with hypothetical profit factors of 1.16 to 1.34. That is too little to separate an edge from noise, and it still trailed buy-and-hold by a wide margin.
The 9.2 on gold looks good on paper, with hypothetical profit factors between 1.5 and 2.5, but it produced only 24 to 41 trades in 12 years. That is too few to tell skill from luck: three or four lucky trades decide the result.
How we tested
- Platform: MetaTrader 5 Strategy Tester, on hourly CFD prices from a retail broker, with daily bars built from the hourly data.
- Period: 1 June 2014 to 1 October 2026, split into 2014–2021 and 2022–2026 to see whether results hold in unseen years.
- Costs: the broker's spread and commission, plus overnight financing on long positions of 6.12% a year for indices and 5% a year for gold.
- Sizing: one position of fixed size, no leverage and no compounding, so results are a percentage of the position value and read the same at any account size. Real CFD accounts use leverage, which multiplies both gains and drawdowns.
- Buy-and-hold: the change in the daily closing price over the same dates, without financing costs. The setup pays financing and buy-and-hold does not, so the comparison is slightly harsher on the setup than holding a CFD would be; the gap is far too large for financing to close it.
- Data: every number on this page is in a free JSON file you can reuse with a link.
What this means for traders
Popular setups often look convincing on a chart because the eye remembers the clean trend days and forgets the many small losses. Measured over 12 years and three markets, with costs included, the 9.1 behaves like a slow, expensive way of being long the market roughly 30% to 50% of the time. Anyone who still trades it is relying on a discretionary entry idea, not a tested system, and the next drawdown could be deeper than the 20–30% seen in this hypothetical backtest. Results on other brokers will differ, and past results never guarantee future ones.