RSI Reversion

Buy oversold, sell overbought. Works in ranges, gets run over by trends.

What it is

Reference — Glowwiki: RSI
The definitions and formulas below follow that article. It states its own limit plainly:
“Whether this produces profit is not addressed in that article.”
That is the question this page answers.

Note from the reference: Wilder smoothing has no fixed window, so an RSI value depends on how much history was loaded before it — Cutler documented that the same symbol and setting gives different values depending where the data file starts. Cutler's variant, using a simple moving average, does not have this dependency. The reference also notes that its Cardwell section has carried an unsourced tag since June 2014, and that Wilder and Cardwell read divergence in opposite directions.

Formulas

RSI

RSI = 100 − 100/(1 + RS), RS = avgGain / avgLoss

Averages use Wilder smoothing (α = 1/n), not a fixed window

Established

RSI is a momentum indicator measuring the speed of price change, conventionally read as overbought above 70 and oversold below 30. [1]

These rest on a single line of sourcing. The Glowwiki reference article draws on the same encyclopaedia cited here, so two domains do not make two independent accounts. Nothing below has been checked against a second, separately grounded source. It is listed as established because the sources state it as definition rather than as finding — not because it has been cross-verified.

Conventional reading

The 70/30 thresholds are conventional rather than absolute and are applied differently across markets and periods. [1]
Levels such as 80/20 or 90/10 occur more rarely but indicate stronger momentum — which is why the threshold is swept rather than fixed on this page. [1, 3]
Wilder and Cardwell read divergence in opposite directions — one as an approaching reversal, the other as trend confirmation. The reference presents both without ruling, and so does this page. [3]
The reference notes its own Cardwell section has carried an unsourced tag since June 2014. The trend-adjusted 40–80 / 60–20 bands attributed to him are therefore weaker ground than the rest of the article, and are not used here. [3]
Marek and Šedivá (2017) reported that over longer horizons RSI trading is usually overcome by simple buy and hold. [3]
The usefulness of technical analysis as a whole is contested by the efficient-market hypothesis. The reference article states the dispute without ruling on it — and neither does this page. What is below is one measurement, not a verdict. [1, 2]

Listed as convention, not as established fact.

Why your chart may not match Why the history matters, and how far back: with Wilder smoothing the weight of a bar k periods ago is (1 − 1/n)^k. At n = 14 that is 0.108 at 30 bars back, 0.0117 at 60, and 0.000137 at 120 — a half-life of about 9.4 periods. This is arithmetic that follows from the definition, not a figure reported by a source. It explains the measurement below rather than replacing it.

Wilder smoothing has no fixed window — every earlier bar still influences today's value, so the same symbol and the same 14-period setting give different RSI readings depending on how much history was loaded. Measured on this data at one point (2023-06-24): 30 bars of history shifts the value by 0.67 against the full-history figure, and at 120 bars that particular gap closes. That is one measurement, not a general bound — across the 365-bar rolling windows used here the mean absolute difference is 0.30 points but the largest is 20.48, because a small weight multiplied by a large price move is not small. The RSI-below-30 entry test flips on 0.22% of bars. This site computes the indicator inside each window rather than carrying history across it.

Sources

RSI — Glowwiki (en)
Technical analysis — Wikipedia (en)
Relative strength index — Wikipedia (en)

The short answer

Beat buy & hold in none of the 3 markets tested.

Gold (PAXG) -101% · BTC/USDT -485%. A rule that wins in one market and loses in another did not find an edge; it found a rising asset. Each figure is measured only against buying and holding that same market over that same period.

Market	Return	Buy & hold	vs B&H	Beat it	Max DD	Underwater	Grid beat
BTC/USDT	45%	530%	-485%	no	66%	2872d	0%
ETH/USDT	-11%	276%	-287%	no	78%	1477d	6%
Gold (PAXG)	17%	119%	-101%	no	17%	493d	0%

All figures on this page are from daily candles unless a row says otherwise.

Evidence

Everything below is the same rules run over real price history. Pick a market to see the fills, the equity path and the parameter grid for it.

Daily candlesBTC/USDT · 3188 bars · 2018-01-01 → 2026-09-23

Total return

45.3%

10,000 → 14,530

vs buy & hold

-484.6%

over the same period 529.9%

Max drawdown

65.8%

10,000 → 5,670

Longest underwater

2872

2872 days · 98% of the period

Settings that beat it

0 / 16

of 16 parameter settings tested

Win rate

69%

95% CI 42%–87% · n=13

The win rate on this page is not a readable number. With 13 closed trades the 95% confidence interval runs 42%–87% — it spans a coin flip, so the headline 69% cannot be distinguished from chance at this sample size. It is shown because leaving it out would be worse, not because it supports a conclusion.

A 0% risk of ruin here is a property of the test, not a safety rating. These runs are long-only spot with no borrowed money, so there is no margin call and nothing can force the position closed. The account cannot be wiped out — it gets stuck holding the asset instead. Capital deployed (2×) and max drawdown (65.8%) are the numbers that carry the risk on this page. The same rules run on margin would be a different measurement, and this figure would not stay at zero.

Daily candles

account equitybuy & holddashed = starting capital

26 fills. Signals are taken on the close and filled at the next open.

Timeframe

Everything above uses the daily row. These are the same rules and the same parameters on a different candle size.

Timeframe	Return	Buy & hold	vs B&H	Beat it	Trades
4H	-41%	524%	-565%	no	87
1D ←	45%	530%	-485%	no	13
1W	102%	524%	-422%	no	2

Outcome distribution

A single return is one chosen start date. These are 95 runs of the same rules with the start shifted forward 30 days at a time.

Worst	p10	p25	Median	p75	p90	Best	Losing starts
-53.2%	-31.2%	-15.9%	4.1%	22.0%	59.6%	132.2%	42%

Does it survive its own parameters?

The headline above uses one parameter setting. Here is every setting tested: 16 combinations, median 45.3%, best 249.5%, worst -46.0%. 0 of them beat buy & hold.

A strategy that only works at one setting did not work — it was fitted.

Parameter grid

RSI Reversion run across its own parameter range over the full period. The same rules flip from profit to loss depending only on the numbers you picked — which is why a single headline return is meaningless on its own.

	buyBelow 20	buyBelow 25	buyBelow 30	buyBelow 35
sellAbove 60	65%	54%	120%	-36%
sellAbove 65	108%	112%	250%	2%
sellAbove 70	70%	36%	45%	-45%
sellAbove 80	55%	-46%	-31%	-42%

Hover a cell for trade count and drawdown. “·” = invalid combination (fast ≥ slow).

Year by year

Year	2018	2019	2020	2021	2022	2023	2024	2025	2026
Return	-31%	-15%	82%	-8%	-40%	49%	14%	3%	-12%
Max DD	64%	30%	13%	32%	50%	9%	21%	19%	29%
Trades	1	0	1	1	1	2	1	1	1

Recent closed trades 13 closed trades — showing the last 13

Entered	Entry	Exited	Exit	Held	Size	P&L (USDT)	Return
2026-02-01	78,780.47	2026-08-20	69,300.11	200 bars	16,550.51	-2,020.76	-12.2%
2025-11-18	92,261.25	2026-01-15	96,903.3	58 bars	15,789.23	+761.29	+4.8%
2025-02-26	88,724.73	2025-05-03	96,838.69	66 bars	14,495.23	+1,293.99	+8.9%
2024-06-25	60,323.45	2024-10-21	68,997.48	118 bars	12,698.33	+1,796.9	+14.1%
2023-08-18	26,636.72	2023-10-21	29,654.22	64 bars	11,429.04	+1,269.3	+11.1%
2023-03-10	20,372.39	2023-03-18	27,381.43	8 bars	8,520.48	+2,908.56	+34.1%
2022-05-10	30,089.27	2023-01-12	17,934.29	247 bars	14,323.86	-5,803.38	-40.5%
2022-01-08	41,587.26	2022-03-30	47,411.07	81 bars	12,589.52	+1,734.34	+13.8%
2021-05-18	43,559.79	2021-07-29	39,999.56	72 bars	13,737.52	-1,148	-8.4%
2020-03-13	4,802.41	2020-04-30	8,774.19	48 bars	7,534.06	+6,203.46	+82.3%
2019-09-25	8,501.8	2020-01-15	8,810.23	112 bars	7,284.86	+249.2	+3.4%
2018-11-15	5,920.16	2019-02-19	3,895.4	96 bars	11,093.56	-3,808.7	-34.3%
2018-02-06	6,943.1	2018-07-24	7,717.79	168 bars	10,000	+1,093.56	+10.9%

Prices include slippage; size is what was actually committed to that trade including fees.

Assumptions

Period	2018-01-01 → 2026-09-23
Parameters	{"length":14,"buyBelow":30,"sellAbove":70}
Rolling test	365 bars × 95 start dates
Input hash	e013d4415f596913

Daily candlesETH/USDT · 3188 bars · 2018-01-01 → 2026-09-23

Total return

-11.3%

10,000 → 8,868

vs buy & hold

-287.1%

over the same period 275.8%

Max drawdown

77.9%

10,000 → 2,593

Longest underwater

1477

1477 days · 97% of the period

Settings that beat it

1 / 16

of 16 parameter settings tested

Win rate

67%

95% CI 39%–86% · n=12

The win rate on this page is not a readable number. With 12 closed trades the 95% confidence interval runs 39%–86% — it spans a coin flip, so the headline 67% cannot be distinguished from chance at this sample size. It is shown because leaving it out would be worse, not because it supports a conclusion.

A 0% risk of ruin here is a property of the test, not a safety rating. These runs are long-only spot with no borrowed money, so there is no margin call and nothing can force the position closed. The account cannot be wiped out — it gets stuck holding the asset instead. Capital deployed (1×) and max drawdown (77.9%) are the numbers that carry the risk on this page. The same rules run on margin would be a different measurement, and this figure would not stay at zero.

Daily candles

account equitybuy & holddashed = starting capital

24 fills. Signals are taken on the close and filled at the next open.

Timeframe

Everything above uses the daily row. These are the same rules and the same parameters on a different candle size.

Timeframe	Return	Buy & hold	vs B&H	Beat it	Trades
4H	-90%	271%	-361%	no	6
1D ←	-11%	276%	-287%	no	12
1W	128%	271%	-143%	no	2

Outcome distribution

A single return is one chosen start date. These are 95 runs of the same rules with the start shifted forward 30 days at a time.

Worst	p10	p25	Median	p75	p90	Best	Losing starts
-73.6%	-47.7%	-13.2%	0.0%	19.1%	99.5%	173.0%	41%

Does it survive its own parameters?

The headline above uses one parameter setting. Here is every setting tested: 16 combinations, median -11.3%, best 461.6%, worst -56.3%. 1 of them beat buy & hold.

Parameter grid

	buyBelow 20	buyBelow 25	buyBelow 30	buyBelow 35
sellAbove 60	131%	256%	-34%	-56%
sellAbove 65	65%	233%	-37%	-41%
sellAbove 70	104%	462%	-11%	-19%
sellAbove 80	46%	66%	-55%	-18%

Hover a cell for trade count and drawdown. “·” = invalid combination (fast ≥ slow).

Year by year

Year	2018	2019	2020	2021	2022	2023	2024	2025	2026
Return	-59%	-5%	99%	0%	1%	5%	2%	-11%	-8%
Max DD	78%	28%	17%	0%	58%	11%	37%	41%	36%
Trades	1	1	1	0	2	1	1	1	1

Recent closed trades 12 closed trades — showing the last 12

Entered	Entry	Exited	Exit	Held	Size	P&L (USDT)	Return
2025-11-05	3,288.69	2026-08-20	2,251.68	288 bars	12,978.45	-4,110.19	-31.7%
2025-03-01	2,238.71	2025-05-09	2,206.29	69 bars	13,195.53	-217.09	-1.7%
2024-07-05	3,061.23	2024-11-10	3,124.64	128 bars	12,953.64	+241.89	+1.9%
2023-08-18	1,682.33	2023-10-24	1,764.59	67 bars	12,374.53	+579.11	+4.7%
2022-11-10	1,103.28	2023-01-12	1,388.71	63 bars	9,850.85	+2,523.69	+25.6%
2022-05-10	2,230.11	2022-08-13	1,957.3	95 bars	11,246.35	-1,395.51	-12.4%
2022-01-08	3,200.28	2022-03-28	3,294	79 bars	10,948.24	+298.11	+2.7%
2020-03-13	107.72	2020-04-30	215.29	48 bars	5,489.04	+5,459.2	+99.5%
2019-11-23	149.62	2020-01-15	165.52	53 bars	4,971.93	+517.11	+10.4%
2019-07-17	199.04	2019-09-20	220.15	65 bars	4,504.17	+467.76	+10.4%
2018-08-08	378.39	2019-02-19	145.62	195 bars	11,727.6	-7,223.43	-61.6%
2018-03-18	548.25	2018-04-24	644.26	37 bars	10,000	+1,727.6	+17.3%

Prices include slippage; size is what was actually committed to that trade including fees.

Assumptions

Period	2018-01-01 → 2026-09-23
Parameters	{"length":14,"buyBelow":30,"sellAbove":70}
Rolling test	365 bars × 95 start dates
Input hash	47546f923545199c

Daily candlesGold (PAXG) · 2214 bars · 2020-09-01 → 2026-09-23

Total return

17.3%

10,000 → 11,730

vs buy & hold

-101.3%

over the same period 118.6%

Max drawdown

17.4%

10,000 → 9,395

Longest underwater

493

493 days · 97% of the period

Settings that beat it

0 / 16

of 16 parameter settings tested

Win rate

75%

95% CI 41%–93% · n=8

The win rate on this page is not a readable number. With 8 closed trades the 95% confidence interval runs 41%–93% — it spans a coin flip, so the headline 75% cannot be distinguished from chance at this sample size. It is shown because leaving it out would be worse, not because it supports a conclusion.

A 0% risk of ruin here is a property of the test, not a safety rating. These runs are long-only spot with no borrowed money, so there is no margin call and nothing can force the position closed. The account cannot be wiped out — it gets stuck holding the asset instead. Capital deployed (1×) and max drawdown (17.4%) are the numbers that carry the risk on this page. The same rules run on margin would be a different measurement, and this figure would not stay at zero.

Daily candles

account equitybuy & holddashed = starting capital

16 fills. Signals are taken on the close and filled at the next open.

Timeframe

Everything above uses the daily row. These are the same rules and the same parameters on a different candle size.

Timeframe	Return	Buy & hold	vs B&H	Beat it	Trades
4H	1%	119%	-118%	no	44
1D ←	17%	119%	-101%	no	8
1W	0%	123%	-123%	no	0

Outcome distribution

A single return is one chosen start date. These are 62 runs of the same rules with the start shifted forward 30 days at a time.

Worst	p10	p25	Median	p75	p90	Best	Losing starts
-11.1%	-5.0%	0.0%	3.2%	7.9%	9.2%	15.3%	23%

Does it survive its own parameters?

The headline above uses one parameter setting. Here is every setting tested: 16 combinations, median 4.3%, best 32.7%, worst -5.2%. 0 of them beat buy & hold.

A strategy that only works at one setting did not work — it was fitted.

Parameter grid

	buyBelow 20	buyBelow 25	buyBelow 30	buyBelow 35
sellAbove 60	0%	12%	18%	4%
sellAbove 65	0%	-0%	0%	-5%
sellAbove 70	0%	12%	17%	11%
sellAbove 80	0%	22%	33%	32%

Hover a cell for trade count and drawdown. “·” = invalid combination (fast ≥ slow).

Year by year

Year	2021	2022	2023	2024	2025	2026
Return	9%	-5%	8%	3%	0%	-6%
Max DD	6%	14%	7%	5%	0%	17%
Trades	2	1	2	0	0	1

Recent closed trades 8 closed trades — showing the last 8

Entered	Entry	Exited	Exit	Held	Size	P&L (USDT)	Return
2026-03-20	4,654.72	2026-08-11	4,394.93	144 bars	12,447.84	-718.21	-5.8%
2024-11-13	2,560.28	2025-01-26	2,767.62	74 bars	11,538.36	+909.48	+7.9%
2023-06-06	1,920.96	2023-10-20	1,962.02	136 bars	11,319.52	+218.84	+1.9%
2023-02-14	1,807.9	2023-03-14	1,921.04	28 bars	10,674.2	+645.31	+6.0%
2022-05-11	1,835.92	2022-11-11	1,747.13	184 bars	11,239.14	-564.93	-5.0%
2021-06-18	1,791.76	2021-11-11	1,851.07	146 bars	10,900.76	+338.38	+3.1%
2021-02-27	1,750.55	2021-05-09	1,861.86	71 bars	10,269.62	+631.14	+6.2%
2020-09-23	1,910.95	2021-01-03	1,966.41	102 bars	10,000	+269.62	+2.7%

Prices include slippage; size is what was actually committed to that trade including fees.

Assumptions

Period	2020-09-01 → 2026-09-23
Parameters	{"length":14,"buyBelow":30,"sellAbove":70}
Rolling test	365 bars × 62 start dates
Input hash	4adb3b9674b8a9d8

Other strategies

Buy & Hold Martingale Grid DCA MA Cross EMA Cross MACD Cross Bollinger Reversion Donchian Breakout Supertrend Stochastic Cross Parabolic SAR