Methodology & risk note: Backtested event study across SPY, QQQ, IWM, and DIA daily price data, 1993-2026, 33,792 strategy variants, frictionless. Results are hypothetical and not investment advice. Past patterns don't guarantee future results. Full method and disclaimer below.
TL;DR
- The setup matters more than the indicator. Of 33,792 tested configurations, 89.7% of long-side variants made money against 8.8% of short-side. Which way you trade moves the outcome more than which oscillator you pick.
- Buying dips beat chasing breakouts. Mean-reversion variants were profitable 55.7% of the time versus 42.9% for breakout, and on the long side mean-reversion led on risk-adjusted edge for all three oscillators.
- Williams %R is the most robust. Long mean-reversion Williams %R is the only row in the long-side oscillator-by-style table where 100% of variants (all 3,456) made money, and its best reliable strategy posted the smoothest ride of the three (CAR/MaxDD 0.57 over 198 trades).
- RSI owns the single sharpest edge, but it's thin. RSI's best reliable variant hit an R-expectancy of 1.16, the highest in the study, on only 51 trades, a small sample that earns a skeptical eye. Stochastic carried the highest median net profit of the three families.
- Holding period barely mattered. Across time exits of 0, 5, 10, and 15 bars, the median R-expectancy sat near zero. Letting the indicator decide the exit had a slight edge.
R-expectancy is the average profit per dollar risked, a risk-adjusted measure of edge; 0.2 or higher is a common "worth trading" bar. CAR/MaxDD is compound annual return divided by the worst drawdown, where higher means a smoother ride.
You've been arguing about the wrong thing
You switched oscillators after a losing streak, sure the next one was the fix. RSI let you down, so you tried Stochastic. Stochastic whipsawed you, so someone in a forum swore Williams %R was cleaner. You are not alone, and you are not lazy. You're just tuning the wrong knob.
Here's the number that should sting. Across 33,792 tested configurations, the choice between RSI, Stochastic, and Williams %R barely moved the outcome. What moved it was direction: 89.7% of long-side variants made money and only 8.8% of short-side did. The oscillator you spent months agonizing over is the paint color. The direction and style you traded it in are the house. People keep repainting a house with no foundation.
So we settled the argument with data instead of opinion. I spent years myself convinced the perfect indicator was one setting away, so this test is personal. We took the three most popular momentum oscillators and ran 33,792 strategy variants across four major US index ETFs over more than three decades of daily history. Long and short. Buying oversold dips (mean-reversion) and chasing momentum (breakout). A spread of indicator settings and holding periods. Every combination scored the same way.
RSI, the Stochastic oscillator, and Williams %R are not far apart, because they all measure the same thing: how stretched price is relative to its recent range. What separates a winning configuration from a losing one is the setup: the direction you trade and the style you trade it in, not the indicator. Get the setup right and all three capture a similar edge. Get it wrong and none of them save you.
How we tested: 33,792 backtests across four ETFs
Here is exactly what's behind every number below.
- Instruments: SPY, QQQ, IWM, and DIA, four major US index ETFs.
- Window: each ETF's full daily history through 2026. SPY runs from 1993, DIA from 1998, QQQ from 1999, and IWM from 2000.
- Variants tested: 33,792 total. RSI 13,824, Williams %R 13,824, Stochastic 6,144.
- What we varied: the oscillator (RSI / Stochastic / Williams %R), lookback and threshold, direction (long / short), and entry style (mean-reversion buys oversold and exits overbought; breakout does the opposite).
- Time exit: a protective cap at 0, 5, 10, or 15 bars. 0 means the indicator alone decides the exit.
- Position rules: $10,000 fixed per position, no compounding, one position at a time.
- Fills: signals are computed on the close and filled at the next open, so there's no look-ahead. The system can't act on information it wouldn't have had in real time.
- Costs: frictionless. No commission, no slippage. The numbers are indicative base rates, not what you'd net after trading costs.
This is an event study, not a finished trading system. It measures the base rates of each configuration so you can see where the edge lives. Across all 33,792 variants, 49.3% were profitable and 14.6% cleared an R-expectancy of 0.2. Those middling averages hide everything interesting, because the spread between the good setups and the bad ones is enormous. That spread is the story.
Finding #1: Does direction beat the indicator?
Yes, and it isn't close. The single biggest divider in the entire study isn't RSI versus Stochastic versus Williams %R. It's long versus short.
| Trade direction | Variants profitable | Median R-expectancy | Median win rate |
|---|---|---|---|
| Long | 89.7% | 0.13 | 59.3% |
| Short | 8.8% | -0.15 | 40.0% |
Nearly nine in ten long-side variants made money. Fewer than one in ten short-side variants did. This gap holds across all three oscillators and dwarfs any difference between the indicators themselves. If you only knew one fact about a configuration before trading it, you'd want its direction, not its oscillator.
The reason is structural. US equity index ETFs drift upward over time and mean-revert on short horizons: they tend to bounce after dips. An oscillator that buys oversold readings and rides the bounce leans with both forces; a short strategy fights both. Indicator tuning did not rescue the short side of a trade positioned against a 30-year tailwind. So the first lesson is blunt: on index ETFs, these oscillators are long tools. It's the same dip-buying edge we walk through step by step in the RSI deep dive on trading the S&P 500 with mean reversion.
Finding #2: Does mean-reversion beat breakout?
On the long side, yes. The second divider is entry style. Across all variants, mean-reversion was profitable 55.7% of the time with a median R-expectancy of 0.03, while breakout was profitable 42.9% of the time with a median R-expectancy of -0.03. Buying the dip beat chasing the move.
That edge is clearest where it counts, on the long side. Here's the head-to-head, long only, by oscillator and style.
| Oscillator | Style | % of variants profitable | Median R-expectancy | Median net profit | Best net profit |
|---|---|---|---|---|---|
| Williams %R | Mean-reversion | 100.0% | 0.18 | $21,869 | $44,715 |
| Stochastic | Mean-reversion | 99.8% | 0.20 | $16,659 | $37,608 |
| RSI | Mean-reversion | 89.6% | 0.20 | $9,021 | $36,926 |
| RSI | Breakout | 93.0% | 0.08 | $8,540 | $34,917 |
| Stochastic | Breakout | 89.1% | 0.08 | $6,704 | $27,251 |
| Williams %R | Breakout | 72.2% | 0.02 | $2,303 | $16,069 |
Read down the median R-expectancy column. Every mean-reversion-long row sits at 0.18 to 0.20; every breakout-long row sits at 0.02 to 0.08. The risk-adjusted edge is at least twice as rich when you buy oversold instead of chasing strength, and it holds for all three oscillators.
One honest exception is worth flagging. RSI breakout-long had a slightly higher share of profitable variants (93.0%) than RSI mean-reversion-long (89.6%). But the edge per trade and the median dollar profit both still favored mean-reversion, so the broad pattern holds even where the profitability count wobbles. On the short side, every single cell was poor: between 0.0% and 27.8% of variants profitable, with negative median edges across the board. Mean-reversion-short was the least-bad of a bad set (Williams %R 27.8% profitable, median R-expectancy -0.02). Breakout-short ran from 0.0% to 4.9% profitable, with median edges of -0.24 to -0.26. None of that is an edge.
Why mean-reversion wins: it's the drawdown
The percentages above tell you mean-reversion wins more often. They don't tell you what the ride looks like. That shows up in two numbers: how often you win, and how deep the worst loss gets.
| Long-side style | Median win rate | Median max drawdown | Median profit factor |
|---|---|---|---|
| Mean-reversion | 68.8% | 20.8% | 1.59 |
| Breakout | 49.5% | 59.6% | 1.10 |
Two completely different trading experiences hide behind the headline. Long mean-reversion's median win rate is 68.8% and its median worst drawdown is 20.8%. Breakout is a coin flip (49.5% win rate) with a typical worst drawdown of 59.6%, nearly triple, and a profit factor of 1.10. That 1.10 means breakout barely takes in more than it gives back; mean-reversion's 1.59 is a real cushion.
The median mean-reversion variant won 68.8% of its trades with a 20.8% worst drawdown. The median breakout variant won 49.5% with a 59.6% hole. On index ETFs, dip-buying isn't just more profitable at the median. A 20.8% hole is a different life from a 59.6% one.
RSI vs Stochastic vs Williams %R: the three-way head-to-head
Once you've fixed the setup to long mean-reversion, which oscillator wins?
They each win something different. There is no single champion, and anyone selling you one is overfitting to a single metric.
| Dimension | Winner | The number behind it |
|---|---|---|
| Most robust (long mean-reversion) | Williams %R | 100% of all 3,456 variants profitable; best reliable variant CAR/MaxDD 0.57 over 198 trades |
| Highest single risk-adjusted edge | RSI | R-expectancy 1.16, the highest in the study, but on only 51 trades (thin sample) |
| Highest median net profit | Stochastic | leads the three families on median net profit per variant |
To make "best reliable variant" concrete, here is the top strategy from each family once you require at least 50 trades and rank by R-expectancy. The exact settings are included so the result is reproducible, not a vague claim.
| Oscillator | Best reliable variant (settings) | R-expectancy | Net profit | CAR/MaxDD | Win rate | Trades |
|---|---|---|---|---|---|---|
| RSI | SPY, mean-rev long, length 4, lower 10, upper 70, time-exit 0 | 1.16 | $13,307 | 0.30 | 82.35% | 51 (thin) |
| Stochastic | SPY, breakout, %K length 14, smoothing 3, lower 10, upper 90, time-exit 0 | 0.81 | $25,338 | 0.32 | 49.30% | 71 |
| Williams %R | SPY, mean-rev long, length 7, lower -95, upper -5, time-exit 0 | 0.49 | $29,650 | 0.57 | 74.24% | 198 |
Williams %R is the robustness winner. Long mean-reversion Williams %R was the only row in the long-side oscillator-by-style table where every variant turned a profit, all 3,456 of them. Other 100% slices exist elsewhere (lookback 2 on that ladder, RSI below 30). Its best reliable strategy also posted the smoothest equity curve of the three, a CAR/MaxDD of 0.57, on the deepest sample (198 trades). When you want a result that survives changes in settings and isn't propped up by a lucky handful of trades, Williams %R is the steadiest.
RSI owns the single sharpest edge. An R-expectancy of 1.16 is excellent, well above the 0.2 "worth trading" line. That caveat matters, though: it came from only 51 trades. A 51-trade sample is small enough that a few outcomes could move it meaningfully, so treat RSI's peak as a tantalizing data point, not a promise. The broader RSI mean-reversion family was also the least robust of the three on the long side (89.6% profitable versus 99.8% and 100%), meaning its tail of losing variants was wider. RSI rewards precise tuning and punishes sloppiness more than the other two.
Stochastic is the steady middle. It posted the highest median net profit across the family and stayed nearly as robust as Williams %R on the long side (99.8% of mean-reversion-long variants profitable). No glaring weakness. Notably, its best reliable variant came from the breakout side, which shows that family-level patterns don't hold for every individual strategy.
For the record, the single best dollar result in the whole study was a Williams %R, SPY, mean-reversion-long variant at $44,715 net profit, from the same corner of the map (long, mean-reversion, index ETF) that almost every other strong result came from. The one exception is Stochastic, whose best reliable variant is a breakout.
Is this just curve-fit to SPY? No. It holds in every market
The fastest way to fool yourself in backtesting is to tune a strategy until it shines on one chart, then mistake that polish for an edge. That single failure sinks more retail strategies than any bad indicator does, which is why we treat it as its own discipline in why most traders fail the missing piece called robustness testing. The defense is simple: does the same setup work on markets it was never tuned to? We ran all four ETFs separately, so we can check.
| ETF | % of long mean-reversion variants profitable |
|---|---|
| SPY | 96.8% |
| QQQ | 96.4% |
| DIA | 95.7% |
| IWM | 93.9% |
The edge doesn't live in one ticker. Long mean-reversion was profitable in the mid-90s percent of variants on every ETF: large-cap (SPY), tech (QQQ), small-cap (IWM), and the Dow (DIA). Stronger still, of the 2,112 distinct parameter combinations we tested, 91.5% made money in all four markets at the same time. A setting that works everywhere at once is the opposite of curve-fit.
Broken down by oscillator, the robustness order from the head-to-head holds market by market. Williams %R was 100% profitable in all four ETFs. Stochastic was at or just below 100% everywhere (its low was 99.2% on IWM). RSI was the consistent weak link, 85.4% on IWM up to 92.2% on SPY, never broken but always the one with the widest tail of losing variants. The picture is the same wherever you point it, which is exactly what you want to see before you trust a pattern.
The champions board: the best variant in every family x market cell
One more way to see the pattern travel: take the single highest-net-profit variant from each of the twelve family x market cells and lay them side by side:
| Market | RSI | Stochastic | Williams %R |
|---|---|---|---|
| SPY | $34,249 (MR, len 2) | $35,144 (MR, len 5) | $44,715 (MR, len 2) |
| QQQ | $36,926 (MR, len 2) | $37,608 (MR, len 5) | $43,553 (MR, len 3) |
| IWM | $25,467 (MR, len 2) | $28,719 (MR, len 9) | $28,166 (MR, len 2) |
| DIA | $21,690 (BO, len 2) | $19,990 (MR, len 5) | $28,034 (MR, len 4) |
Twelve champions, twelve family x market maxima, and they all tell the same story. Every one of the twelve is long. Eleven of twelve are mean-reversion (the lone breakout on the board is RSI's DIA champion). Nine of twelve use a lookback of five bars or less, the fast-oscillator rule again. And Williams %R holds the top dollar result on three of the four markets, losing only IWM to Stochastic by $553. When the maxima keep landing in that corner of the parameter space (long, mean-reversion, fast lookback), that corner isn't luck. That's where the edge lives.
How should you actually set one up? Two knobs that matter
Say you've accepted the verdict: long, mean-reversion, on an index ETF. Two settings still decide how well it works: how fast the oscillator is (its lookback length), and how deep an oversold you wait for (its threshold). The data has a clear answer on both.
Faster is more robust
| Lookback (RSI and Williams %R) | % of long mean-reversion variants profitable |
|---|---|
| 2 bars | 100.0% |
| 3 bars | 99.8% |
| 4 bars | 99.1% |
| 5 bars | 95.1% |
| 7 bars | 89.4% |
| 8 bars | 85.2% |
A short lookback makes the oscillator twitchy. A long lookback makes it sluggish. For dip-buying on index ETFs, twitchy wins. The share of profitable variants falls in a clean staircase as you lengthen the lookback: across the RSI and Williams %R families, which share a length setting, a 2-bar lookback was profitable in every mean-reversion-long variant; an 8-bar one in 85.2%. The per-trade edge barely moves (median R-expectancy stayed near 0.18 to 0.19 the whole way), and median net profit actually fell from $22,403 at length 2 to $13,469 at length 8. The answer to "what period should I use?" is blunt: a short one. The classic RSI(14) default sits outside that ladder. Stochastic was swept at 5, 9 and 14, and its best reliable variant by R-expectancy is length 14.
How deep an oversold? Moderate, not extreme
The other knob is the threshold: how oversold the oscillator has to get before you buy. Instinct says deeper is better: wait for a real washout. The data says there's a sweet spot, and the extreme is a trap.
| RSI oversold threshold | Per-trade edge (R-expectancy) | % of variants profitable | Typical trades (33 yrs) |
|---|---|---|---|
| Below 5 | 0.18 | 63.0% | ~6 (too thin) |
| Below 10 | 0.26 | 86.3% | ~34 |
| Below 15 | 0.23 | 95.0% | ~78 |
| Below 20 | 0.21 | 93.6% | ~152 |
| Below 25 | 0.18 | 99.5% | ~222 |
| Below 30 | 0.18 | 100.0% | ~294 |
Read the two columns against each other. The per-trade edge rises as you demand a deeper oversold, but only up to a point. It peaks at "RSI below 10" (R-expectancy 0.26) and then fades. Push to the extreme, "RSI below 5," and the signal fires about six times in thirty-three years, well below the study's 50-trade reliability floor, and only 63.0% of those variants were profitable. Meanwhile reliability climbs the other way: the shallower, more frequent thresholds (below 25, below 30) were profitable in 99-100% of variants and booked the most total dollars, but with a gentler edge per trade.
So there's no single right answer. There's a trade you get to make. Want the sharpest edge per trade and you'll accept fewer signals? Aim moderate-deep, around RSI below 10-15. Want the steadiest, most reliable version with the most trades and the most total profit? Stay shallower, around below 25-30. What you should not do is chase the deepest washout you can find. That's where the sample gets too thin to trust, a trap we cover in more depth in understanding Z-score and its application in mean-reversion strategies.
Does the holding period matter? Barely.
A natural worry: maybe the protective time exit, how long you hold before bailing, is quietly doing the work. It isn't.
| Time exit | Median R-expectancy | Share clearing R-expectancy ≥ 0.2 |
|---|---|---|
| 0 bars (indicator decides) | ~0.00 | 23.4% |
| 5 bars | ~0.00 | ~10-12% |
| 10 bars | ~0.00 | ~10-12% |
| 15 bars | ~0.00 | ~10-12% |
The median edge sits at essentially zero no matter which holding cap you pick. The one visible difference is that letting the indicator decide the exit, with no time cap, produced the most strategies clearing the 0.2 bar: 23.4%, roughly double the capped settings. If anything, cutting winners short with a tight time exit costs you a little. Let the setup play out.
The verdict, and the honest caveats
One finding runs through every section above: the setup matters more than the indicator. All three oscillators capture the same edge: long, mean-reversion, buying oversold dips on these four index ETFs. Short loses. Breakout is weaker. Among the three, Williams %R mean-reversion-long is the most robust, RSI carries the single highest but thin risk-adjusted edge, and Stochastic posts the highest median net profit. "Which oscillator is best?" is the wrong question. "Which setup?" is the right one, and the answer is long mean-reversion.
Now the limits, because a study that hides them isn't worth trusting.
- Frictionless. Every number was computed with no commission and no slippage. Real costs would shrink any edge here, and the thinner ones might not survive at all. Read the magnitudes as relative, not as take-home returns.
- A parameter sweep, not a deployable system. This measures a flat-only sweep, one position at a time, so overlapping signals are skipped. It tells you where edges concentrate. It does not hand you a tuned, risk-managed strategy. That's separate, careful work.
- Thin samples are flagged, not buried. RSI's headline 1.16 rests on 51 trades. The most extreme settings everywhere carry small samples. Lean on the shape of the findings (long beats short, mean-reversion beats breakout, Williams %R is the steadiest) rather than the last decimal of any single cell.
- No short edge here is specific to index ETFs. Short variants were 8.8% profitable on SPY, QQQ, IWM, and DIA. On individual stocks, crypto, or instruments that do not share this sample, the short side could look completely different. Don't generalize the "stay long" ranking beyond these four index ETFs.
- Long beating short is not buy-and-hold. On SPY, buy-and-hold finished at $157,639.52. Williams %R's best reliable variant finished at $29,650. The matched random control on SPY averaged $10,636.55. The ranking inside this sweep is long over short. It is not a claim these variants beat holding the index.
Findings at a glance
| Finding | The number | What it means |
|---|---|---|
| Direction is king | 89.7% of long variants profitable vs 8.8% short | Long variants beat short variants on these four ETFs. The short side loses. |
| Style matters next | Mean-reversion 55.7% profitable vs breakout 42.9% | Buy oversold dips; don't chase breakouts. |
| Drawdown is why | Mean-reversion 20.8% max drawdown vs breakout 59.6% | Dip-buying wins often and shallow; breakout is a coin-flip with a deep hole. |
| Not curve-fit | Edge holds in all 4 ETFs; 91.5% of settings profitable in all four at once | The pattern travels across markets, not tuned to one chart. |
| Faster is sturdier | 2-bar lookback 100% profitable vs 8-bar 85.2% | Use a short lookback; lengths 2 to 4 held up better than 5 to 8 on the RSI and Williams %R ladder. Stochastic was tested at 14. |
| Oversold sweet spot | Edge peaks at RSI below 10; RSI below 5 fires ~6 times in 33 yrs | Buy moderate oversold, not the deepest washout. |
| Williams %R = robustness | 100% of 3,456 long-MR variants profitable | The steadiest of the three across settings. |
| RSI = sharpest edge | R-expectancy 1.16 on 51 trades | The highest peak, but a thin sample. Verify before trusting. |
| Stochastic = highest median profit | leads the three families | A strong, no-glaring-weakness middle option. |
| Holding period | median R-expectancy ~0.00 across 0/5/10/15 bars | The entry drives the edge; let winners run. |
What this means for you
- Fix the setup before you touch the indicator. Decide direction (long) and style (mean-reversion) first. Those two calls decided 33,792 outcomes far more than the choice of RSI, Stochastic, or Williams %R.
- On these four index ETFs, the ranking is long and buy the dip. Shorting made money in 8.8% of variants. The median breakout variant sat through a 59.6% worst drawdown. Buy-and-hold on SPY still finished higher than the reliable oscillator variants in this sweep.
- Set it fast and moderately oversold. A 2-4 bar lookback was far more robust than a 5 to 8 bar one on the RSI and Williams %R ladder. Stochastic was swept at 5, 9 and 14. Aim for a moderate oversold (around RSI below 10-15) for the sharpest edge, or shallower (below 25-30) for the steadiest, most-traded version.
- Pick your oscillator by what you value. Want robustness? Williams %R. Want the sharpest single edge and you'll do the verification? RSI. Want a steady all-rounder with the best median profit? Stochastic.
- Distrust any single dazzling number. RSI's standout 1.16 came from 51 trades. Before you trade a backtest's best cell, check that its neighbors hold up too. The shape of a finding is trustworthy; one lucky row is not.
Methodology: how this data was generated
This is a backtested event study, not live trading results. The numbers come from our own in-house research engine run over the full daily price history of each ETF, not from third-party summaries.
- Data source: Daily OHLCV (open, high, low, close, volume) price data for SPY, QQQ, IWM, and DIA.
- Date range: each ETF's complete history through 2026. SPY from 1993, DIA from 1998, QQQ from 1999, IWM from 2000.
- Variants: 33,792 total (RSI 13,824, Williams %R 13,824, Stochastic 6,144), sweeping oscillator settings, direction (long / short), entry style (mean-reversion / breakout), and time exit (0 / 5 / 10 / 15 bars).
- Entry and exit: signals computed on the close and filled at the next open (no look-ahead). One position at a time, $10,000 fixed per position, no compounding.
- Scoring: each variant measured for percent profitable, median and best net profit, win rate, R-expectancy (average profit per dollar risked), and CAR/MaxDD (compound annual return over worst drawdown). "Reliable" variants require at least 50 trades.
- Frictionless assumption: computed without commissions or slippage. Real-world costs would reduce any edge shown here.
The same short, oversold, index-ETF setup shows up across our indicator work, from the ADX indicator and how to actually make it work in real trading to the cumulative RSI twist on RSI(2) for the S&P 500. For more on why even a clean backtest is not the same as a live edge, watch Why Most Profitable Backtests Fail in Live Trading on the StatOasis YouTube channel.
Disclaimer
All results here are derived from historical backtesting using daily ETF price data and do not represent actual trading results. Backtested performance is hypothetical. Past performance of any pattern does not guarantee future results. These results are for educational and informational purposes only and do not constitute investment advice. StatOasis is not a registered investment advisor. Nothing here is a recommendation to buy or sell any security. Please consult a licensed financial professional before making any investment decision.
Transparency: StatOasis sells trading-education products. Our research is produced independently and is not altered to favor a sale.
Freshness: The data is current through 2026. We re-review these studies when the underlying datasets are extended.
Get the next study in your inbox
This is one study in an ongoing series. If you want the next one, the same kind of large-scale, myth-busting, frictionless-but-honest backtest, join The Overfit newsletter at StatOasis.com/Overfit. It's where we publish the data behind the trading ideas everyone argues about.
The best oscillator isn't a secret you're missing. It's the setup you've been ignoring while you argued about the paint color.
For the developer, paste-ready: the block below is valid JSON-LD (one @graph). Drop it as-is into a single <script type="application/ld+json"> tag in the page <head>. Before publishing, confirm the URL / @id paths resolve (author page /about/ali-casey, logo.png, the /terms license) and update datePublished / dateModified if the live date differs.
