CaseyC% Oscillator: A Mean Reversion Indicator That Trades Different Days From RSI(2), 32 Years of SPY and 20 of ES

TL;DR: the answer box

• Different days. Only 13.5% of CaseyC% entries on SPY, and 11.5% on ES, fell on a day RSI(2) also bought.
• Spread the size. Five rules at one unit each had a worst drop of $36,757 on SPY against $63,085 for RSI(2) at five units. Net profit: $319,768 against $338,592.
• Not the best rule in hindsight. CaseyC% alone at five units scored a Ret/DD of 12.69 on SPY against 8.70 for the five-rule account. On ES the best rule changed from one half of the test to the other.
• The setting decides the score. At the published settings, CaseyC% scored a Ret/DD of 12.74 on SPY against 5.48 for RSI(2). The typical CaseyC% setting scored 5.81, and RSI(2) at its own best level 8.19.
• A stable region exists. With a 3-day change, a buy below 20 and smoothing 3, CaseyC% beat RSI(2) below 10 on both markets at all 8 range lengths tested.

What is CaseyC%, and how is it different from RSI(2)?

A mean reversion trade buys a market that just dropped hard, betting on a bounce. The classic tool for it is RSI(2). RSI, the Relative Strength Index, is a 0 to 100 score that compares recent gains with recent losses. The (2) means it looks back only two days.

CaseyC% starts from a different number. I designed it in 2022. It works in three steps:

1. Take the % change of the close over the last 3 days. That is the plain rate of change indicator.
2. Ask where today's change sits inside its range over the last 11 days. 0 is the bottom of that range. 100 is the top.
3. Smooth the result over 3 days.

So RSI(2) asks "how big were the last two days' losses next to the gains?" CaseyC% asks "how bad is this 3-day drop compared with the last two weeks?" Different questions. Different days.

The smoothing can push it a few points below 0 or above 100. That is part of its definition, not a bug.

How I tested CaseyC% against RSI(2)

I spent years hunting for the best indicator. It is one of the reasons I lost money early on. So this test does not ask which indicator is best. It asks what each one does, and what happens when you run them together.

The five rules. All long only. All sell at the open 4 bars after they buy. No stop.

• RSI(2): buy when RSI(2) closes below 10.
• Regular CaseyC%: buy when CaseyC% closes below 20.
• Cumulative CaseyC%: add up the last 3 CaseyC% readings, and buy when the sum is below 60.
• Dynamic CaseyC%, 5 and 10 bars: buy when that sum is the lowest of the last 5 bars, or of the last 10.

The markets. SPY from 1994-02-01, and E-mini S&P 500 futures (ES) from 2006-01-03, both to 2026-08-20. That is 32.55 years of SPY and 20.63 of ES. SPY trades on $35,000 in whole shares. ES trades one contract. No compounding.

The fills: Execution rule v2. The signal is read at the close. The trade buys the next open. The exit sells at the open 4 bars later: buy Tuesday's open, sell the next Monday's open. You can't buy at a close you have not seen yet, so this study never does.

The sweep. 800 CaseyC% settings per market, plus RSI(2), cumulative and dynamic variations. 1,644 backtests in all.

The scores, in plain words:

• Win rate: the share of trades that made money.
• Drawdown (worst drop): the biggest fall from a high point, in dollars.
• Ret/DD: net profit divided by the worst drop. Profit per dollar of pain. Higher is better.
• Typical setting: the median, the middle one when you line them all up.
• Reliable: a setting with 100 or more trades.

What does one CaseyC% trade look like?

Here is a trade RSI(2) never took. On 207 of the 312 ES CaseyC% trades, RSI(2) was at or above 10 on the signal day, so the RSI(2) rule had no signal. This chart shows the one whose profit sits closest to the middle of those 207. A typical one, not the best.

CaseyC% read 14.6. RSI(2) read 20.2. The trade bought the open of 2011-07-14 and sold the open of 2011-07-20. +9.00 points, $450 on one contract.

How do CaseyC% and RSI(2) compare at the published settings?

No rule wins every column. Here are all five side by side.

On SPY, the gap is the drop. CaseyC% below 20 took 481 trades and won 61.3%, with a worst drop of $5,679 and a Ret/DD of 12.74. RSI(2) below 10 took 427 trades and won 59.3%, with a worst drop of $12,349 and a Ret/DD of 5.48.

On ES, it is closer. RSI(2) made more, $157,712 against $149,875. CaseyC% had the smaller worst drop, $33,512 against $56,450, so its Ret/DD came out higher: 4.47 against 2.79.

Look at the win rates: 61.3% against 59.3%. Almost the same. The risk number is what separates them.

The cumulative and dynamic versions did not lift SPY. They scored 6.70, 5.54 and 5.17, all below regular CaseyC%. On ES the dynamic 5-bar rule scored the highest Ret/DD of the five, 6.43. Check its neighbours before you build on it. By lookback, 3 to 20 bars, it ran 5.79, 6.43, 3.07, 3.96, 0.94 and 1.07 on ES. That is not a smooth hill. It is a jagged ridge.

Is it just the published setting?

Partly, yes. That is the honest part.

The published setting, CaseyC% (3, 11, 3) below 20, came out of my earlier SPY research. So on SPY it is not out of sample: it was not tested on fresh data it had never seen.

On SPY it ranks 14 of 759 reliable settings. On ES it ranks 230 of 726. Near the top on the market it was built on. Middle of the pack on the other one.

Now compare the typical setting at each entry level with RSI(2) at the same level.

On SPY, RSI(2) scored higher than the typical CaseyC% setting at every entry level tested. Across all 759 reliable SPY settings, 54.3% beat RSI(2) below 10. Only 20.2% beat RSI(2) below 20, its best level.

On ES, the picture splits. 69.8% of CaseyC% settings beat RSI(2) below 10, a weak ES setting. Only 0.8%, 6 of 726, beat RSI(2) below 20. And RSI(2) itself jumped around on ES: 2.79 below 10, 9.18 below 20, 2.28 below 30.

So which one is better? Wrong question. Neither one ranks first everywhere. They do different jobs.

Is there a stable region of CaseyC% settings?

Yes. With smoothing 3, every range length passed.

Picture the settings as a map. On a plateau, you nudge a setting and the score barely moves. On a spike, one nudge drops you off a cliff. A spike is what curve-fitting looks like: a rule tuned so tightly to the past that one small nudge breaks it.

Here is the test. Hold the change length at 3 and the entry at below 20. Move the range length and the smoothing. A cell passes if it has 100 or more trades AND a Ret/DD above RSI(2) below 10 on SPY and on ES.

21 of 40 cells passed. The smoothing 3 row passed at all 8 range lengths, from 5 to 21. Inside that row:

• SPY: Ret/DD from 6.28 to 12.74, typical 8.23, on 331 to 675 trades.
• ES: Ret/DD from 3.26 to 5.61, typical 4.45, on 205 to 452 trades.

The published default sits inside that row. It is not a lone peak. I set this test after the sweep, though, so it describes the map. It does not prove the future.

Don't chase the peak. The single best SPY cell, (3, 9, 2) below 10, scored 17.08. I don't use it. It is the top of an 800-cell search. Picking the top of a search is the risk researchers measure as the probability of backtest overfitting. I did not compute that probability here.

And "almost everything makes money" proves little: 797 of 800 CaseyC% settings were profitable before costs on each market. The S&P 500 rose. Pick almost any CaseyC% setting, on either market, and you made something before costs.

Does CaseyC% trade different days from RSI(2)?

Yes. This is what CaseyC% was built for, and it holds up.

Of 481 CaseyC% trades on SPY, only 65 started on the same day as an RSI(2) trade. That is 13.5%. On ES, 11.5%. The cumulative and dynamic versions shared even fewer: as low as 8.0% on SPY.

Now the honest half. Different entry days are not the same as unrelated profits. A correlation measures how much two profit streams move together: 1.00 means as one, 0 means unrelated.

CaseyC% and RSI(2) correlated at 0.54 on SPY and 0.55 on ES. Across all ten pairs, 0.40 to 0.63 on SPY and 0.49 to 0.65 on ES. Moderate. All five buy dips in the same index, and their profits move together. Not twins.

Is it better to run five variants or one variant at five times the size?

Better than piling it onto a typical rule. Not better than the best rule in hindsight.

Say you have room for five units. Do you spread them over five rules at one unit each? Or put all five on one rule?

I built both as accounts. One $175,000 account holds all five rules at one unit each. Against it, one $175,000 account per rule holds that rule at five units. Same money. Same total size. The only difference is the spread.

Against RSI(2) at five units, the rule most dip traders already run, the five-rule account:

• SPY: worst drop $36,757 against $63,085, 41.7% smaller. Net profit $319,768 against $338,592.
• ES: worst drop $121,612 against $282,250, 56.9% smaller. Net profit $777,962 against $788,562.

The worst single year tells the same story. In 2018, RSI(2) at five units lost $19,446 on SPY. The five-rule account lost $7,690.

But it did not beat the best single rule. Rank all six accounts by Ret/DD and the five-rule account comes second on both markets.

On SPY, CaseyC% alone at five units scored 12.69 against 8.70, with a smaller worst drop, $28,430 against $36,757. On ES, the dynamic 5-bar rule edged it, 6.43 against 6.40, with a deeper worst drop, $155,188 against $121,612.

Think of it as a band against a soloist. On the night, the best soloist will outplay the band. You book the band when you don't know in advance who that soloist will be.

Can you pick the winning rule in advance?

Not by picking the past winner on ES. The leader changed between halves.

On ES, the best single account from 2006 to 2016 was cumulative CaseyC%, at 3.43. From 2016 to 2026 it was the dynamic 5-bar rule, at 6.74. That same rule scored 0.33 in the first half. Pick the first-half winner, and you held a rule that came fifth of six in the second half.

The five-rule account ranked 3, 3, 3 and 2 of 6 across the four halves. Never first. Never near the bottom.

On SPY, regular CaseyC% led both halves. That is the case for it, and also the case for caution: its settings came from SPY in the first place.

Costs do not change the five-rule account's Ret/DD lead over RSI(2). With the engine's cost schedule, the five-rule account nets $298,162 on SPY with a Ret/DD of 8.05, against $321,758 and 4.99 for RSI(2) at five units. On ES, $726,806 and 5.86 against $751,152 and 2.63. One ES order flips: after costs, the five-rule account's 5.86 moves past the dynamic 5-bar rule's 5.44.

Do these signals beat a random entry?

On SPY, mostly. On ES, rarely.

A random control buys on random days instead of signal days. It asks for as many entries as the typical CaseyC% setting: 531 on SPY and 345 on ES. One trade at a time, so it takes fewer: 421 on SPY and 274 on ES. It holds 4 bars and sells at the open the same way, averaged over 10 seeds. Random long entries make money on a rising index too: $28,421 on SPY, with a spread of $15,821.

The luck bar is the random average plus two spreads. A setting above it made more than the random average by over two spreads.

Of 291 CaseyC% settings whose trade count is within 25% of the control's requested entries, 59.5% cleared the luck bar on SPY. On ES, also 291 settings, 19.9% did.

The published rule finished 2.8 spreads above the random average on SPY and only 1.4 on ES.

And the context every dip article should print: buy and hold SPY made $518,686 on the same $35,000 over the same years. CaseyC% below 20 made $72,321, in the market only 23.5% of the time. These rules earn far less than holding, with far less time exposed and a far smaller worst drop.

The verdict, and the honest limits

What held up. CaseyC% buys on different days: 86.5% of its SPY entries came on days RSI(2) stayed out. Spreading the size over five dip rules cut the worst drop by 41.7% on SPY and 56.9% on ES against RSI(2) at five units, for similar profit. And with smoothing 3, CaseyC% has a stable region on both markets.

What did not. CaseyC% does not outscore RSI(2) everywhere. The published setting is a top cell on SPY, where it was found, and a middling one on ES. The cumulative and dynamic versions did not lift SPY. And five rules together did not beat the best single rule picked after the fact.

The limits, in plain words:

• Two related markets. SPY and ES both track the S&P 500. Long only, daily bars. Nothing here covers other markets, intraday charts or short trades.
• Not out of sample on SPY. The published settings came from earlier SPY research.
• No costs in the headline numbers. The cost check above trims every account.
• No stop loss. The largest single loss on one ES contract was -$27,275 for CaseyC% and -$16,150 for RSI(2).
• A backtest is not a promise. It is the best evidence about the past.

What this means for you

1. Use CaseyC% for the days it trades, not as a replacement. Most of its entries fell on days RSI(2) did not buy: 86.5% on SPY and 88.5% on ES. Don't pick it, or drop it, on one setting's score.
2. Spread your size. On these markets, five rules at one unit each beat one typical rule at five units on the worst drop.
3. Check the pair, not the parts. Before you combine rules, look at their correlation and their joint drawdown. Here they correlated at 0.40 to 0.65 and still cut the worst drop.
4. Trade the plateau, not the peak. With smoothing 3, every range length passed on both markets. The 17.08 peak is the top of an 800-cell search, so I don't use it.
5. Pair every win rate with a risk number. CaseyC%'s 61.3% win rate on SPY came with a Ret/DD of 12.74. The dynamic 5-bar rule's 57.5% came with 5.54.
6. Get it free. CaseyC% is a free download in the Bonus section of the StatOasis community, so you can run this test on your own markets.

Don't hunt for the best indicator. Hunt for the different one.

Reference tables

For readers who want the detail behind the charts.

Table 1. The five rules, one unit each

Execution rule v2, next-open fills, sell at the open 4 bars after entry. SPY $35,000 in whole shares, ES one contract. Profit factor: dollars won for each dollar lost.

Table 2. Five rules at one unit, or one rule at five units ($175,000 account)

No costs, no compounding. Same total size in every account.

Methodology and honesty footer: 1,644 backtests on SPY (1994 to 2026) and ES (2006 to 2026) daily bars, Execution rule v2 (signal at the close, fill at the next open, time exit at the open 4 bars after entry). Flat-only per rule, SPY $35,000 in whole shares, ES one contract, no compounding. The account test uses one $175,000 account per comparison. Headline figures have no costs, with a costed check. Buy and hold and a seeded, frequency-matched random control computed on the same bars. Every figure here comes from the study's own results. This is research, not investment advice.

Want the next study when it is finished, including the ones that overturn something I believed? That is what the newsletter is for: StatOasis.com/Overfit