Awesome Oscillator Tested: 14,000 Backtests, and 5/34 Ranks 28th

TL;DR: the answer box

• All four signals have an edge, and it is small. On the long side, median return-to-drawdown ran 2.06 for the zero-line cross, 1.65 for the first green bar, 1.26 for the saucer and 1.17 for twin peaks. Every one has a positive edge per trade and a profit factor above 1.00. The spread between the best and worst rule is 0.90.
• Williams' 5/34 ranks 28th of the 35 pairs tested. Median return-to-drawdown 1.30, against 2.34 for the best pair, 13/55. The default is not the best default, and nobody on the first page of Google has checked.
• The exit is worth 1.7 times the signal. The time exit moves median return-to-drawdown by 1.52; the choice of signal moves it by 0.90. The thing the indicator is famous for is not the thing that decides the result.
• The side you trade dwarfs everything. 92.0% of long variants made money against 8.6% of short ones, a gap of 2.33 in median return-to-drawdown, 2.6 times the gap between the best and worst signal.
• None of it beat holding the index, and the median setup is indistinguishable from a coin flip. Of 1,215 reliable long variants on SPY, zero beat buy-and-hold's 8.82% a year. The median one returned 1.38% a year at 0.05 per unit of worst drawdown, against 0.051 for a seeded random control that traded just as often. What the Awesome Oscillator did do is fall less far: 27.32% at worst against the index's 56.47%.

What is the Awesome Oscillator, and what is the formula?

It is a subtraction, and the whole of it fits on one line.

Median Price = (High + Low) / 2
AO           = SMA(Median Price, 5) - SMA(Median Price, 34)

Both averages are simple, never exponential. Both are taken on the median price, the midpoint of each bar's range, and never on the close. Those two details are the entire identity of the indicator, and they are the two things most write-ups skip.

The result is drawn as a histogram around a zero line. A bar is coloured green when it is higher than the bar before it and red when it is lower, which is a statement about the change in momentum rather than its direction. That is also how TradingView documents the indicator, and how MetaTrader's own reference describes it. Bill Williams introduced it in New Trading Dimensions in 1998, alongside the three signals this study tests.

The three signals Williams published, in his own terms:

• Zero-line cross. The histogram crosses from below zero to above it. Momentum has turned positive.
• Saucer. Above the zero line, two consecutive bars fall and the third turns back up. A pause inside a move that is already running.
• Twin peaks. Below the zero line, the oscillator makes a trough, rallies without crossing zero, then makes a second, higher trough and turns up off it.

And the fourth rule, which is not Williams' but is what the tutorials show:

• First green bar. The histogram turns up after at least one bar of turning down. Buy the colour change.

I tested all four because a study that only tested the canon would be answering a question nobody types into a search box.

How is the Awesome Oscillator different from the MACD?

This is the most-asked question about the indicator and the most commonly fumbled answer. The two differ in two ways, not one.

Most explanations mention the periods and stop. But a 5/34 MACD is still built on closes with exponential weighting, and StockCharts' own MACD reference is explicit about both. So it is not an Awesome Oscillator. It is a different measurement that happens to be plotted the same way. The median price is what makes the AO a statement about where the market traded rather than where it settled.

How we tested

One harness. The only thing allowed to vary is the signal.

• Markets: E-mini S&P 500 futures (ES, regular session, 5,026 daily bars from 2007-01-03) plus SPY (8,398 bars from 1993-02-02), QQQ (6,866), IWM (6,537) and DIA (7,153). 135 years of daily bars pooled.
• Variants: 14,000, being four signal rules × 35 fast/slow pairs × long and short × two exit rules × five time exits × five markets. 2,800 per market.
• Entry: the rule fires on a close, the position is filled at the next open. Exit: either the mirror-image rule fires against the position, or the oscillator crosses zero against it, or a time exit of 0, 3, 5, 10 or 20 bars fires first. Zero means no time exit at all.
• Money: $35,000 a position, no compounding, one position at a time. Frictionless in the main sweep, and priced properly further down.
• No look-ahead, structurally. Every rule reads only bars up to and including the one it fires on. Twin peaks confirms one bar after its second trough, because a trough is not visible until the following bar closes above it. A test in the repo truncates the price series at three points and fails the build if any rule changes its mind about a past bar once later bars exist.
• Reliability: 11,943 of the 14,000 variants clear a 50-trade floor. Thin cells are flagged, never quietly dropped, and the long side of one whole rule lives mostly below that floor (see the twin-peaks section).

One thing is held still, and it is the indicator's own definition: simple averages on the median price. Sweeping the averaging method too would have tested a different indicator, and would have let anyone argue the result was decided by the smoothing rather than by the signal.

Do the Awesome Oscillator's signals work at all?

Yes, all four of them, and by less than you would hope.

Every rule clears a profit factor of 1.00 and every one has a positive edge per trade. That is not nothing, and it is not proof of skill either. Randomly timed long entries over the same history also made money, at 0.051 per unit of worst drawdown on SPY, which is the comparison the control section below makes.

Note what the win rate does here, because it is the trap this whole site exists to point at. Twin peaks has the highest win rate of the four at 60.2%, and the worst return-to-drawdown. It wins more often and earns less, because it wins small and loses big. A win rate quoted without a risk-adjusted number beside it is a marketing figure, not a result.

Which matters more: the signal, or the side you trade?

Not close.

Every rule was mirrored bar for bar on the short side, sign-flipped and not re-tuned, so no rule was handed an advantage the other side never got. The long side's median return-to-drawdown is 1.57. The short side's is −0.76. That gap of 2.33 is 2.6 times the 0.90 spread between the best and worst signal.

This is the same finding the Casey Bands study landed on with a completely different indicator family, and the Z-score mean-reversion work before it. On US equity indices, an indicator that tells you to be long is mostly telling you what the index already did for thirty years.

What actually decides an Awesome Oscillator result?

I can rank it, because every axis was swept identically and the answer falls out of the same table.

The time exit is worth 1.7 times the choice of signal. And the direction of that effect is the useful part: holding until the rule says otherwise scores 2.55, and forcing an exit after three bars drops it to 1.03. These are slow signals on daily bars. Cutting them short at three bars more than halves that score.

Every one of those "how to use the Awesome Oscillator" pages spends its length on which pattern to look for. The pattern is the third most important decision on the list.

What are the best Awesome Oscillator settings?

Not 5/34. This is the finding I expected least and checked hardest.

Sweeping the fast average across 2, 3, 5, 8, 13 and 21 and the slow across 21, 26, 34, 44, 55 and 89 gives 35 valid pairs. Williams' published 5/34 scores a median return-to-drawdown of 1.30 and ranks 28th of 35. The best pair, 13/55, scores 2.34, an 80% improvement.

Now read the size of it before you change anything. The full range across all 35 pairs is 1.49. The gap between trading long and trading short is 2.33. Tuning the periods is a smaller decision than picking a side, and no larger than choosing how you exit. If you go and re-optimise your Awesome Oscillator to 13/55 and keep everything else the same, you have spent your effort on the fourth most important thing on the list.

I spent years doing exactly that. Nudging an indicator's periods, watching the backtest improve, calling it progress. What it actually taught me is that a number you tuned until it looked good tells you nothing you did not already put there.

The honest read of that heatmap is not "use 13/55". It is that the whole surface is flat and slightly positive, which is what a weak, real effect looks like. No single pair stands out as a spike above its neighbours, which is the shape a tuned setting leaves.

How often does twin peaks actually happen?

This is the section that changed my mind about the whole indicator.

At the published 5/34 setting, on the long side:

Twin peaks fires roughly twice a year. Only 780 of 1,750 long twin-peaks variants reach even 50 trades across a full market history, and 50 trades is the bare floor at which a win rate starts to mean anything.

That is a statement about sample size, not about performance. But it has a hard consequence: any published twin-peaks win rate that does not tell you how many occurrences it rests on is unfalsifiable. Not wrong. Unfalsifiable, which is worse. On one market you would need most of a working lifetime to accumulate enough of them to judge. This study pools five markets and 135 years of daily bars and still only finds 330 of them.

Does any of it beat just holding the index?

No. Not one variant, on return.

A note on the drawdown number, because two definitions circulate and only one is what you mean. The figure below is the largest percentage fall in account equity, which is the one a person pictures when they hear "max drawdown". Held against SPY from 1993, that is 56.47%, dated 2009-03-09.

Of the 1,215 reliable long variants on SPY, 0.0% beat buy-and-hold on return. The best earner in the entire sweep made 4.48% a year against the index's 8.82%, roughly half the money, while being in the market 85% of the time.

The pattern holds on every market tested:

But look at the second column of the first table before writing the indicator off. The median Awesome Oscillator variant fell 27.32% at its worst where the index fell 56.47%, while sitting in the market less than a third of the time. 5.3% of variants beat the index on return per unit of drawdown.

The Awesome Oscillator is a smoke alarm, not a compass. It is very good at telling you to leave the building. It has never once known where you were going.

Is it better than a coin flip?

Barely, and for the median setup, no.

Every study here is scored against a seeded random control: entries scattered at random through the same price history, firing as often as this study's own median variant and holding for as long, averaged over ten seeds. The only difference between the control and the strategy is where the entries fall.

On SPY the random control returns 0.051 per unit of worst drawdown. The median reliable Awesome Oscillator variant returns 0.05.

Those are the same number. The median long Awesome Oscillator setup on SPY performs like randomly timed entries, though the control's frequency and hold are matched to every reliable SPY variant rather than to the long ones alone, and the platform default of 5/34 ranks 28th of 35 pairs, below that median. Rules above that median exist. This study does not count how many of them clear the control's seed-to-seed spread, and clearing the mean is not clearing the spread, so nothing here says any single rule has beaten luck.

This is why the frequency section matters and why the parameter surface matters. A weak effect that is genuinely there will be flat, broad and hard to beat luck with. That is exactly what this looks like.

What happens when the trading isn't free?

Almost nothing, and that is not the reassurance it sounds like.

Taking the best SPY variant of each rule and charging both sides of every trade against its own notional:

Each of those four best variants keeps most of its profit, because they trade between 69 and 245 times across three decades rather than thousands of times. On the Casey Bands study costs reversed the ranking outright, because those variants traded over a thousand times. Here they barely move it.

So on these four best SPY variants, costs are not the problem: each kept between 87% and 92% of its frictionless profit at 0.05% a side. The absence of enough return is.

Note also that all four of those rows are hand-picked winners: the single best variant of each rule on one market. They are printed so you can see what the ceiling looks like, not because they are tradable. That is what the next section is for.

Where the results actually hold up

A single top row is the least reliable number in any sweep. The useful question is which configurations were profitable in every market, at a trade count worth taking seriously.

614 of 1,019 full configurations were profitable in all five markets with a median of at least 100 trades, 60% of them. By rule: the first green bar in 298 of 344, the zero-line cross in 184 of 306, the saucer in 122 of 302, and twin peaks in 10 of 67.

The stable region's own median return-to-drawdown is 1.92, against 5.31 for the best configuration inside it. The distance between those two numbers is the size of the error a single cherry-picked backtest makes. If you take one number off this page, take 1.92, not 5.31.

Two things stand out in that list. Every top row uses a fast period of 13 or 5, not the published 5 with a 34, and 38% of the all-five-markets configurations run a fast period of 8 or longer, slower than Williams'. And the rule that dominates the durable list is the first green bar, which is the one Williams never published.

The verdict, and the honest limits

The Awesome Oscillator is a real indicator with a small, real, defensive edge, sold as something it is not. The alarm works. It was never a compass.

What is true:

• All four of its signals have a positive edge per trade and a profit factor above 1.00 on the long side, across five markets and 135 years of pooled daily bars.
• The median reliable SPY variant's worst drawdown is much smaller, 27.32% against buy-and-hold's 56.47%, and it is out of the market 74% of the time. The two were measured side by side, not causally.
• The four best SPY variants keep between 87% and 92% of their profit at 0.05% a side, because these rules trade rarely.

What is not:

• It does not out-earn holding the index. Zero of the 1,215 reliable long SPY variants beat buy-and-hold's 8.82% a year on return. On return per unit of worst drawdown, 5.3% of them did.
• Its published 5/34 setting is not its best setting, and ranks 28th of 35.
• Its most-taught pattern, twin peaks, fires about twice a year, which is not enough to evaluate on one market in a working lifetime.
• The median setup performs like a coin flip trading at the same rate.

And the limits of this study, stated plainly:

• Five US equity index markets, daily bars. They are highly correlated. Nothing here says anything about currencies, commodities, single stocks or a five-minute chart, and the settings do not carry over untested.
• Nothing here isolates a mechanism. The long-versus-short gap says where the edge sits. It does not say what produces it. No test here separates the oscillator's timing from the direction it was pointed in.
• Long-only conclusions rest on a thirty-year bull market. The direction finding is real, but it is a fact about US indices in this period, not a law.
• Frictionless in the main sweep, with costs priced separately rather than baked in. The costs section is the honest version.
• No stop losses. Every exit here is a signal or a bar count. Stop losses sit outside the exit rules this study tested, so nothing on this page measures what one would do to any of these numbers, drawdown included.

What this means for you

If you are using the Awesome Oscillator to find trades on these five US equity index markets, the numbers do not back it. None of the 1,215 reliable long SPY variants beat buy-and-hold on return, and 5.3% of them beat it on return per unit of worst drawdown. The signal is also the third most important choice you make, behind the side you take and the way you exit. Fixing the first two is worth more than any pattern on the chart.

If you want the smaller drawdown the median reliable variant showed, the honest framing is that you are testing a defensive filter, not an entry signal. Test it as one. Compare it to the simplest possible alternative, being out of the market when price is below its own long average, before concluding the oscillator earned its complexity.

And if what actually interests you is the long-side edge the numbers keep pointing at, the next thing to test is a rule that buys dips directly. 36 Ways to Buy the Dip is thirty-six versions of exactly that trade, run through this same harness with the same benchmarks and the same refusal to quote a win rate on its own. It is the book I would hand someone who read this page and asked what to do instead.

Whatever you run, run it against buy-and-hold and against a random control before you trust it. That is a habit, not a tool. Though if you want the monitoring side automated once a rule is live, AlgoChef is what I built for it.

An indicator that cannot beat doing nothing is not a strategy. It is a reason to do nothing, which is worth something too.

Published 29 November 2024 · rebuilt from scratch and last updated 21 September 2026. Every figure above is computed from this study's own results table. This is research, not investment advice, and a backtest is not a promise. Trading involves risk of loss. Every number here is frictionless unless a cost is stated, and past results do not predict future returns.

Want the next one? Every study here goes out to the Overfit list first: one measured result a week, no hype, no signals for sale. Join it at StatOasis.com/Overfit.