RSI in crypto measures the balance of recent upward and downward price changes under a specified calculation. An overbought or oversold reading describes that calculation; it does not instruct you to sell or buy. Interpret RSI only after checking the market, candle interval, completion state, smoothing method and the price behavior you are trying to investigate.
A reading of 75 can mean sustained upward momentum. It can also sit beside a trade plan whose downside is unacceptable. The number alone cannot choose between those interpretations. Here is a hand-checkable example that follows one RSI reading through a changing candle and a different calculation method.
What the indicator actually contains
The familiar RSI formula is 100 − 100 ÷ (1 + average gain ÷ average loss). Gains and losses are derived from changes in the selected price series. The oscillator runs from zero to 100, subject to the implementation's treatment of edge cases. Fidelity describes the conventional 70 and 30 reference levels and cautions that readings can remain extreme during strong trends. Fidelity's RSI indicator guide.
A period setting counts bars, not necessarily days. Fourteen one-hour changes and fourteen daily changes summarize different horizons. They also require a starting observation: fourteen changes need fifteen closes for the initial simple-average calculation used below.
State which price enters the calculation. A closing-price RSI cannot be reproduced from a screenshot of candle highs. Venue, pair and spot-versus-contract choice can change the underlying prices. If an AI summary omits these fields, its RSI claim remains incomplete.
The market data freshness guide provides a record for these inputs. A correct formula applied to the wrong or unfinished series is still the wrong evidence for a completed-candle rule.
Reproduce a 14-period seed by hand
Take a fictional asset with an initial close of 100 quote units. Repeat the changes +3, −1 seven times. There are fourteen changes and fifteen closes. The final close is 114 because seven gains of three contribute 21 and seven losses of one subtract seven.
For this explicitly chosen initialization, average gain is 21 ÷ 14 = 1.5, and average loss is 7 ÷ 14 = 0.5. Relative strength is 1.5 ÷ 0.5 = 3. RSI is 100 − 100 ÷ 4 = 75.
Nothing in this construction assigns a probability to the next movement. It simply produces a series and a reproducible indicator value. It also shows why “75% bullish” is an incorrect conversion: an oscillator level of 75 is not a forecast success rate. The confidence-score audit explains how to evaluate numbers that do claim to be probabilities.
Keep full precision until display. If a platform rounds intermediate averages, or initializes them from a different amount of history, its displayed reading can differ. Compare inputs and initialization before declaring one platform wrong.
One developing candle, two very different readings
Continue the same constructed series. During the next candle, price rises from the previous close of 114 to 118. Later, the candle closes at 107. A live indicator can change substantially between those observations.
Using Wilder-style smoothing with period 14, update each average as (previous average × 13 + current gain or loss) ÷ 14. TradingView documents its RSI calculation using a smoothed moving average; the exact implementation and initialization still need to match the platform being checked. TradingView RSI documentation.
| Observation | Gain for this bar | Loss for this bar | Updated gain average | Updated loss average | RSI |
|---|---|---|---|---|---|
| Developing close at 118 | 4 | 0 | 23.5 ÷ 14 | 6.5 ÷ 14 | 78.33 |
| Final close at 107 | 0 | 7 | 19.5 ÷ 14 | 13.5 ÷ 14 | 59.09 |
These are two alternative updates from the same previous completed state, not two consecutive smoothing steps. At 118, RSI is 100 × 23.5 ÷ (23.5 + 6.5) = 78.33. At the final close of 107, it is 100 × 19.5 ÷ (19.5 + 13.5) = 59.09.
A rule requiring a completed RSI above 70 never qualified on that candle. An intrabar rule might have qualified earlier, but it is a different rule and requires its own execution and evaluation assumptions. Do not retrospectively describe an intrabar observation as a completed signal.
This is often called repainting in casual discussion, but the useful diagnosis is more specific: was the indicator legitimately updating on an unfinished bar, or did the implementation use information unavailable at the recorded decision time? The multiple-timeframe worksheet applies the same timing discipline across different intervals.
Why another calculator may show 56.25
Now compute a simple rolling average over the latest fourteen changes after the final −7. The oldest +3 leaves the window. Remaining gains total 18; losses total 14. RSI becomes 100 × 18 ÷ (18 + 14) = 56.25.
That is not the 59.09 Wilder-smoothed result because the calculation changed. Wilder smoothing carries prior averaged information forward; a simple rolling window discards one old change at each step. Specify the method before comparing screenshots or asking AI to explain the disagreement.
TA-Lib's public RSI implementation provides an inspectable example of Wilder-style gain and loss updates and initialization choices. It is code for a particular library, not evidence that every trading application uses identical behavior. TA-Lib RSI source.
Zero-loss and flat-series cases deserve explicit tests. Dividing by zero needs a defined branch. A platform may return a boundary value, a missing value or another documented convention for a flat series. If the method is not known, preserve the uncertainty rather than assigning a universal answer to every implementation.
Add context without assembling a vote count
Start with a question about price. Are you investigating whether recent gains have persisted, whether a sharp move has slowed, or whether a proposed entry is far from an invalidation level? RSI can help describe those observations. It cannot answer all of them by itself.
Pairing RSI with another price-derived indicator does not automatically create independent confirmation. RSI and MACD both transform price history. Three bullish labels can repeat the same information in different forms. Count the underlying evidence, not the number of colored badges.
Volume can add a different measurement, but only with a defined venue, unit and comparable period. The volume confirmation article shows why a volume ratio can change when its denominator or time boundary changes. It does not turn volume into a guarantee either.
Divergence needs a predeclared comparison. Name the two price pivots, the corresponding RSI observations and when each pivot became knowable. Selecting attractive pivots after the subsequent move creates a story that may not have existed at the time.
Use scale changes as a calculation check
The worksheet also offers a simple diagnostic. Multiply every fictional close by ten, preserving the same sequence and method. Gains and losses both multiply by ten: the seed averages become 15 and five, their ratio stays three, and the seed RSI remains 75. The later readings likewise remain unchanged under this proportional rescaling.
Adding the same constant to every close also preserves the changes in this worksheet. It therefore preserves this RSI calculation. These are properties of the specified arithmetic, not evidence about actual market behavior. They can help detect a spreadsheet that accidentally mixes absolute prices with changes.
Neither property makes RSI scores from unrelated assets equivalent forecasts. Two instruments can share an oscillator level while having different liquidity, spread, volatility and loss scenarios. Use the check to validate implementation, then return to the specific market question. A calculation that passes a synthetic invariant can still be unsuitable for a trading decision.
Turn the observation into an explicit research question
A reproducible research note could say: “For this specific spot series, evaluate whether a completed hourly RSI crosses above 30 after previously being below it.” Then define the exact crossing convention, eligible times, costs, entry rule and exit rule separately.
That sentence is a test specification, not a recommended strategy. Compare it with a simple baseline using data that did not determine the rule. Include all tested thresholds and variants in the research record. Choosing 28 rather than 30 after inspecting outcomes is another research choice that must be disclosed.
Check the position math independently. An RSI threshold does not determine an acceptable order size, stop distance or account exposure. A stop can fill worse than its trigger, and a limit order can remain unfilled. If the trade plan lacks a coherent loss scenario, the indicator does not repair it.
A useful RSI note is short and auditable
Keep the exact market, interval, close timestamps, completion flag, input field, period, smoothing and initialization with the displayed value. Add the hypothesis being investigated and what would invalidate it. Save enough source observations to reproduce the calculation.
KLINEO's AI analysis supports reviewing market ideas. This article does not claim that a particular KLINEO screen uses the calculation choices in the worksheet. Inspect the actual data and implementation before transferring an example into an account action.
This article was prepared with AI assistance under the editorial policy. Its price series is fictional, arithmetic is explanatory, and neither RSI nor an AI interpretation of RSI establishes an investment edge or suitability for a reader.