Which Crypto Indicators Actually Work? Testing the Popular Ones

Introduction

Take 23,602 completed daily bars across eight major crypto assets. A randomly chosen day was followed by a higher price one week later 49.9% of the time. That number is the benchmark almost no indicator write-up publishes, and without it a 55% hit rate cannot be told apart from luck. This article runs eight standard signals against it: RSI at both thresholds, both MACD crossovers, both Bollinger band touches, and the golden and death crosses. Nothing was optimised and no setting was searched. The results are uncomfortable in places. One of the most widely taught entry signals in crypto came in below a coin flip. The "overbought" reading was the best bullish condition in the sample, and the death cross outperformed the golden cross on every seven-day measure.

What Is the Base Rate an Indicator Has to Beat?

Any signal must be judged against what would have happened anyway. Take 23,602 completed daily bars from eight major crypto assets. A randomly chosen day was followed by a higher price seven days later exactly 49.9% of the time. That is the number to beat.

The number every signal must beat

The unconditional forward return is the honest comparison, and it is rarely published alongside indicator claims. Measured across the full sample, the next day closed higher 50.3% of the time. The next week managed 49.9% and the next month 50.0% (Binance Spot API, 2026-09-10) . A signal with a 55% hit rate sounds impressive in isolation. Against a 49.9% base it is worth roughly five percentage points, which is a much smaller claim and the only one the data supports. Any test that omits this comparison cannot distinguish a working indicator from the market's own tendency to rise. That is the single most common flaw in published indicator results.

Why the median matters more

The averages tell a different story from the hit rates, and the gap is the most important structural fact in the sample. Over 30 days the mean return was +8.65% while the median was 0.00%. Half of all one-month windows finished flat or lower, and the entire average came from a right tail of enormous rallies. That skew means an indicator can raise the average return while lowering the odds of any individual trade working. It also means reporting only a mean, which most indicator marketing does, hides the shape of the outcome completely. A reader sees one number and assumes it describes a typical outcome, when in this sample the typical outcome was zero.

Eight rules make the comparison concrete.

How Were the Eight Indicator Signals Tested?

Eight standard signals were coded as unambiguous rules and run across eight assets and 23,602 daily bars. Each was scored on what happened over the following one, seven and thirty days. Nothing was optimised, no parameter was searched, and every setting is the textbook default.

Eight signals, one rule each

The rules use the textbook settings rather than tuned ones, because tuning is how a backtest starts describing the past instead of testing a claim. RSI runs at 14 periods with the conventional 30 and 70 thresholds. MACD uses 12, 26 and 9. Bollinger Bands use 20 periods at two standard deviations. The moving-average pair is 50 and 200. Each rule fires on the daily close, and the outcome is measured from that close forward. This article does not re-explain how the indicators are computed. CoinPaprika's dedicated guides to RSI, MACD and Bollinger Bands cover the mechanics. A companion article on technical analysis covers why several of these tools overlap so heavily ↗ in the first place.

Data current as of September 2026.

What counts as a result

A result is the percentage change in the close from the signal day to the day one, seven or thirty sessions later. No stop, no target, no position sizing. That is deliberately crude. It isolates the informational content of the signal from every discretionary decision layered on top of it in real trading. The eight assets are bitcoin, ether, solana, BNB, XRP, cardano, dogecoin and chainlink. The earliest bar is 17 August 2017 (Binance Spot API, 2026-09-10) . Every figure below comes from that one pull, so the signals are all measured on the same bars over the same period. Nothing was excluded after the results were seen.

The most familiar buy signal is the natural place to start.

Does an Oversold RSI Actually Mark a Bottom?

RSI below 30 did beat the base rate on frequency. It closed higher a week later 59.7% of the time against a base of 49.9%, which is a real and repeatable improvement. Over thirty days, though, the same signal returned less than doing nothing at all.

Oversold paid more often

Across 889 fires, an RSI reading under 30 produced a mean seven-day return of +2.46% and a median of +2.11%. Some 59.7% of instances finished positive (Binance Spot API, 2026-09-10) . Against a 49.9% base and a +1.65% base mean, that is a genuine improvement on both counts. The median matters here more than usual. It sits close to the mean, which means the result is not being carried by a handful of violent bounces. Most oversold readings were followed by a modest recovery, which is what the textbook claims and, unusually for this test, what the data shows. Of the eight signals, this is the one whose conventional description survives contact with the sample most cleanly.

But not more in total

The thirty-day picture reverses. RSI under 30 returned a mean of +7.37% over the following month, below the +8.65% unconditional average, despite winning 56.9% of the time. Frequent small gains, in other words, at the cost of missing the rallies that produce the base rate's average. Buying oversold conditions systematically avoided the biggest up-moves in the sample, because those moves start from strength rather than weakness. The signal is real at a one-week horizon and actively counterproductive at a one-month one. That is a sharper horizon dependence than most indicator write-ups acknowledge. The same rule can be defended or attacked depending purely on the window chosen to report it.

Its mirror image behaved even less like the textbook.

Is an Overbought Reading a Reason to Sell?

RSI above 70 is conventionally read as a warning to take profit before a move exhausts itself. In this sample it was the strongest bullish condition tested, by a wide margin. The mean thirty-day return was +21.13% against a base of +8.65%, with 61.3% of instances finishing positive.

Overbought was not a sell

Across 1,929 fires, an RSI reading over 70 produced a +6.26% mean over seven days and +21.13% over thirty. The thirty-day median was +8.61% and the hit rate 61.3% (Binance Spot API, 2026-09-10) . Every one of those figures beats the base rate comfortably. Selling on overbought readings would have exited exactly the periods that produced the sample's returns. The upper Bollinger band told the same story from a different formula. Closes above it returned +18.09% over thirty days with a 56.4% hit rate, against the same +8.65% base. Two separately constructed measures of stretched price agreed.

Momentum beats mean reversion here

The explanation is that crypto's daily series is strongly trend-persistent at these horizons. An asset at a stretched momentum reading is usually in the middle of a move rather than at the end of one. Mean-reversion language survives from markets that behave differently. It gets applied to crypto by inheritance rather than by evidence, and this sample gives that inheritance no support at these horizons. That does not make overbought a buy signal in any tradeable sense. Entering after a large move carries drawdown risk that a hit rate does not capture. It does make the standard reading of the indicator backwards for this sample.

One signal underperformed a coin flip outright.

Which Signal Performed Worst Against the Base Rate?

The MACD bullish crossover is one of the most widely taught entry signals in crypto. It closed higher a week later 49.1% of the time, against a base rate of 49.9%. It did worse than the market's own tendency, and it did so in every cut of the data tested.

The crossover that missed

Across 825 fires, the MACD line crossing above its signal line produced a seven-day mean of +1.40% and a median of -0.20%. The hit rate was 49.1% (Binance Spot API, 2026-09-10) . The negative median is the telling number: more than half of these signals were followed by a lower price a week later. At one day the hit rate was 48.4%. The bearish crossover, read conventionally as a sell, was followed by a 54.8% one-day hit rate, which is again the opposite of the textbook. Both crossovers are lagging by construction, since a crossover can only occur after the move that caused it. By the time the lines cross, the information that moved them is already several sessions old.

Data current as of September 2026.

Below the base rate everywhere

A single weak result could be noise. This one was not. The MACD bullish crossover finished below the base-rate hit rate in every cut of the data tested later in this article: 47.5% when overlapping signals were removed, 51.7% on bitcoin alone against a 53.2% bitcoin base, and 46.3% from 2023 onward. Four independent slices, four results below the comparison. That consistency is stronger evidence than any single number, and it points the same way each time.

The famous moving-average signals produced the sharpest surprise.

Did the Golden Cross Beat the Death Cross?

It did not, and the gap was not small. The death cross, universally reported as bearish, was followed by a +3.26% mean seven-day return and a 65.7% hit rate. The golden cross, its bullish twin, returned +2.26% with a 48.5% hit rate over the same window.

The death cross outperformed

Across 70 fires each, the bearish crossover of the 50-day below the 200-day average beat its bullish twin on mean, median and hit rate at seven days. It led on hit rate at thirty days too, 61.4% against 55.4% (Binance Spot API, 2026-09-10) . The golden cross's seven-day median was negative at -0.41%, meaning most golden crosses were followed by a lower price a week later. Both signals require roughly 200 days of history before they can fire, so both are describing something that finished months ago. The death cross appears to mark capitulation more reliably than the golden cross marks the start of a recovery. One fires into fear and the other into a move that has already been running for months.

Seventy events is not many

The honest caveat is the sample. Seventy fires across eight assets and nine years is a thin basis for a strong claim, and a handful of events could move these numbers materially. On bitcoin alone the crossovers fired fewer than ten times each, too few to report at all. The direction of the result held in the de-clustered and 2023-onward cuts, which is reassuring but not conclusive. Treat this finding as a well-supported reason to doubt the conventional reading rather than as a tradeable edge in its own right.

Counting matters as much as measuring.

Why Do Raw Backtest Numbers Overstate a Signal?

RSI stayed below 30 for days at a time, so 889 raw fires represent only 158 distinct episodes. Counting each day as its own signal inflates the sample more than fivefold. It makes a modest number of events look like real statistical weight, and it flatters almost every threshold rule.

Counting the same signal twice

A threshold rule is not an event. It is a state, and a state persists. When RSI drops under 30 and stays there for six sessions, the raw count records six signals whose forward windows overlap almost entirely. They are one observation recorded six times. Re-run the test with a minimum thirty-day gap between fires of the same signal on the same asset. RSI below 30 collapses from 889 to 158, RSI above 70 from 1,929 to 239, and the upper Bollinger touch from 1,839 to 381 (Binance Spot API, 2026-09-10) . Only 12.4% of RSI-above-70 fires survive that rule, against 58.4% of MACD crossovers and 98.6% of death crosses. A crossover genuinely is a discrete event rather than a persistent state, and that difference in kind matters more than any difference in formula.

What de-clustering costs

The results shrink with the sample. RSI below 30 falls from a 59.7% seven-day hit rate to 53.8%, and RSI above 70 from 56.1% to 54.8%. Both still beat the base rate, but by roughly half as much. The MACD bullish crossover gets worse, dropping to 47.5%. The death cross barely moves, holding 65.2% across 69 independent episodes, which is the strongest sign that its result is not an artefact of overlapping windows. Any indicator claim quoting thousands of signals from a threshold rule is quoting overlapping observations. The number of independent events is the one that matters.

Subsets are the next test.

Do the Results Hold on Bitcoin and in Recent Years?

Splitting the sample two ways tests whether a result depends on one asset or one era. Bitcoin alone and the period from 2023 onward both broadly reproduce the full-sample findings. The exceptions are as instructive as the confirmations, and they fall on the signals with the thinnest samples.

Bitcoin only, and 2023 onward

Bitcoin's 3,311 daily bars carry a friendlier base rate than the pooled sample. Its seven-day hit rate is 53.2% with a +0.53% median, against the pool's 49.9% and -0.01%. Against that tougher comparison, RSI below 30 hit 58.0% and RSI above 70 hit 64.0%, both still ahead. The MACD bullish crossover managed 51.7%, again below its own base. From 2023 onward, across 10,776 asset-days, RSI below 30 hit 61.6% and the lower Bollinger touch 58.8%. The MACD bullish crossover fell further, to 46.3%, its weakest showing in any cut (Binance Spot API, 2026-09-10) .

Data current as of September 2026.

What survived every cut

Three findings held in all four views of the data. The MACD bullish crossover underperformed its base rate every time. An oversold RSI beat its base rate every time, though the margin varied from four to twelve points. The death cross outperformed the golden cross wherever both fired often enough to compare. Everything else moved enough between cuts to be treated as unstable. That is a modest list to draw from eight signals and 23,602 bars, and the modesty is itself the finding. Three stable results out of eight tested rules is closer to what the evidence supports than any confident ranking of indicators.

Even the survivors come with limits.

What Does This Test Not Prove?

A forward-return test measures whether a signal carries information about what comes next. It does not measure whether trading on that signal makes money. The distance between those two questions is wide, and it is where most published strategies quietly fail.

What this test cannot settle

No fees, spreads or slippage are included, and at eight assets with hundreds of fires those costs would consume much of a five-point edge. A companion article measures what execution actually costs on a live book ↗. Nothing here accounts for drawdown between entry and the measurement date. A signal that finishes higher after thirty days may have been unbearable to hold on day ten. The test also uses a single fixed horizon rather than an exit rule, and real strategies exit on conditions rather than on a calendar. Each of those omissions flatters the results rather than harming them, which is the direction that matters. The measured edges are therefore ceilings rather than estimates.

Costs, sizing and exits

Position sizing is absent entirely. These assets differ enormously in volatility, and an equal-weight reading treats a dogecoin signal as equivalent to a bitcoin one. Survivorship is a smaller issue here than in most crypto backtests, since all eight assets still trade. That is itself a selection, though: tokens that failed are not in the sample. A companion article covers the full catalogue of backtesting pitfalls ↗. The honest summary is that these numbers set an upper bound on what the signals could deliver, and every real-world consideration lowers it.

That leaves a short reading routine.

How Should You Judge Any Crypto Indicator Claim?

Five questions separate a testable indicator claim from marketing copy. None of them requires statistical training, and none requires access to the underlying data. Most published claims fail at the very first question, and almost none survive all five in sequence.

Reading any indicator claim

Ask for the base rate before anything else, because a hit rate without one is meaningless. Ask how many independent events sit behind the number rather than how many signal days. Ask which horizon the result applies to, since this test found signals that worked at one week and inverted at one month. Ask whether the result survives a different asset or a different period. Then ask whether costs are included. A five-point edge does not survive many round trips at crypto spreads. A claim that answers all five is rare enough to be worth reading carefully. A claim that answers none of them is not evidence about anything.

Data current as of September 2026.

A short checklist

The measured conclusion from this test is narrow and worth stating plainly. Two of eight signals beat their base rate consistently, one underperformed it consistently, and the rest moved around enough to be unreliable. That is not an argument against technical analysis, which a companion article frames as a description of the present rather than a forecasting engine. It is an argument against treating any single threshold rule as an edge without checking it first. The tools are cheap to test, the data is free, and the base rate is the only benchmark that matters.

FAQ

References / Sources