What trading indicators measure, and what they cannot

A trading indicator is a formula. It takes a series of past prices, and for some indicators past volumes, and returns another series; nothing in the arithmetic knows anything that its inputs did not. An indicator is therefore only as good a description of the past as its inputs are. In crypto those inputs are the weak part: on a centralised venue every price and every volume figure starts as a trade record that the venue keeps about itself, and a data site’s chart combines many such records.

The indicators discussed here are decades older than Bitcoin. The Relative Strength Index and Average True Range come from J. Welles Wilder’s 1978 book, on-balance volume from Joe Granville’s 1963 one, and MACD from Gerald Appel in the late 1970s. What follows states each construction plainly, says what it summarises, and then looks at what the crypto market does to the numbers that go in.

An indicator is a summary, not a signal

A moving average is price with the short-term movement averaged out. The simple version is the average of the last N closing prices; the exponential version weights recent prices more, applying a multiplier of 2 / (N + 1) to the newest value. StockCharts’ guide to both names the cost of the smoothing: an average of past data trails the price, and trails it further the longer the window.

The Relative Strength Index compares recent gains with recent losses. It first appeared in Wilder’s New Concepts in Technical Trading Systems, and StockCharts and TradingView describe the same construction. Take the closing prices and treat each period’s change as a gain or a loss. Average the gains and the losses over 14 periods, the default in both sources. Divide the average gain by the average loss to get RS, and set RSI to 100 − 100 / (1 + RS), which confines it between 0 and 100. After the first 14 periods the averages are smoothed rather than recomputed: each new average is the previous one multiplied by 13, plus the latest gain or loss, divided by 14.

MACD is the difference between a 12-period and a 26-period exponential average. A nine-period exponential average of that difference is the signal line, and the histogram is the gap between the two lines. StockCharts’ guide gives the same three parts.

Every term in these formulas comes from the past, and none refers to anything else. An average, a ratio of averages or a difference of averages cannot contain what its inputs did not, so a reading is a statement about the window it was computed over. An RSI of 70, for instance, is algebra: inverting the formula gives RS = 70 / 30, which says the recent average gain was about 2.3 times the recent average loss. The 70 and 30 lines that make a chart look like an alarm are an interpretation laid over the number. TradingView’s documentation attributes them to Wilder, as the levels above which a market is to be considered overbought and below which oversold. “Overbought” is a word with an expectation inside it, and the formula never states that expectation.

The parameters are settings. The sources above give 14 for RSI and 12, 26 and 9 for MACD, and each is counted in candles, not in time. A 14-period RSI is a statement about 14 candles, so what it spans depends on how long the chart’s candles are. That point returns below.

What “volume” is on a crypto venue

Volume is the quantity traded over an interval, and what it means depends on who counted. On a centralised crypto venue the count is the venue’s own: its matching engine records the trades, and its published totals and data feeds are what everyone else sees. Decentralised exchanges record trades on a chain, which is a different case and not covered here.

Data sites then collect those figures. CoinGecko’s methodology, as read on 3 October 2026, says it collects price, volume and liquidity data across all tickers on the exchanges it has integrated. An exchange’s total is the sum of the volumes of its trading pairs, leaving out pairs that it has blacklisted for inconsistent data and that have not updated for over three hours. Its API documentation shows the unit being collected: a ticker is one pair on one venue, with a base currency, a target currency, a last price and a volume. The volume on a CoinGecko page is therefore that site’s total of figures from the venues it integrates, after its own filters. The exchange reserves table on this site holds a different figure, the on-chain balances DefiLlama attributes to each exchange, and no volume.

US stock markets take a different route. Under Rule 603(b) of Regulation NMS, every exchange on which a stock covered by the national market system (an NMS stock) trades, and the national securities association, must act jointly under national market system plans to disseminate consolidated information on quotations for and transactions in those stocks. The SEC’s 2020 market data release describes the core data the existing plans carried: the price, size and exchange of the last sale, each exchange’s best bid and offer, and the national best bid and offer across all of them. The same release moved the rule towards competing consolidators in place of a single processor, but what comes out is still consolidated data: the trades in each stock, assembled across venues under a common rule. Alongside it sits an old prohibition. Section 9(a)(1) of the Securities Exchange Act of 1934 makes it unlawful to effect a transaction involving no change in beneficial ownership, in a security other than a government security, in order to create a false or misleading appearance of active trading.

That does not make every printed trade in a US stock genuine. It means a stock has a consolidated volume figure assembled by rule across venues, and a legal prohibition on the specific practice that inflates volume.

The EU’s MiCA regulation has a counterpart to the prohibition. Article 91 bans market manipulation, and Article 86 applies that ban to crypto-assets admitted to trading or for which admission has been requested. Article 91 defines manipulation to include entering a transaction or placing an order that, unless carried out for legitimate reasons, gives or is likely to give false or misleading signals as to the supply of, demand for or price of a crypto-asset (paragraph 2, point (a)(i)).

What differs is the reporting, which under MiCA is per venue and not combined. Article 76 requires a crypto-asset service provider operating a trading platform to publish the price, volume and time of each transaction on that platform as close to real time as is technically possible (paragraph 10). It must also make that data available free of charge in machine-readable form 15 minutes after publication, and keep it published for at least two years (paragraph 11). A search of the regulation’s published text for this piece found no mention of a consolidated tape or a reference price, and the word “consolidated” appears only in the sense of consolidated financial statements.

A volume figure, whoever assembles it, does not say how much of itself carried risk. A trade between two accounts under one owner is counted exactly like a trade between strangers. The wash-trading mechanism is set out in its own piece, along with the best-documented measurement of it: a 2019 Bitwise presentation, filed in the SEC’s comment record, that judged about 4.5 per cent of the bitcoin volume traded against fiat currencies or stablecoins, as reported across the top venues one aggregator listed, to be real. That was one week in March 2019, presented by an applicant making a case. Even when every print is real, volume is the wrong measure of liquidity for a second reason: it records what traded, not what could have, and the second is what a trade of size pays for, as depth versus volume explains.

Some indicators take volume as an input. On-balance volume, in the construction StockCharts describes, adds the whole period’s volume to a running total when the close is above the previous close, subtracts it when the close is below, and leaves the total unchanged when the two are equal. If a period’s volume includes trades that moved nothing, the running total moves by that amount all the same.

Fragmentation and the missing reference price

Price has the same problem in another form. A US stock has a national best bid and offer, a reference computed across exchanges under a common rule. A crypto asset trades on many venues at once, each with its own order book and its own last price, so a chart that shows “the price” has either chosen one venue’s series or built a blend.

CoinGecko publishes how it builds its blend. For each coin it starts from the 600 highest-volume tickers, removes those whose prices are outliers, and takes the volume-weighted average price of the rest. That is a published method with a dependency: which tickers count, and how much each weighs, are set by reported volume. A price built that way has the volume figure folded into it, so an indicator that uses only price is not clear of a volume figure it never displays.

The unit being blended is also less uniform than it looks. On 3 October 2026 the first page of CoinGecko’s public tickers endpoint for bitcoin, ordered by volume, returned 100 tickers, 97 of which had bitcoin as the base asset. Counting those 97 for this piece, they came from 72 venues, 19 of them listing more than one quote asset, and were quoted in nine different assets: 58 against USDT, 19 against USDC, 8 against US dollars, 5 against euros, and 7 against five others, among them Korean won and Japanese yen. That is a sample of the tickers with the highest reported volume, not a census of markets. USDT and USDC are tokens designed to track the dollar, not dollars, and a token’s price against the dollar is a market of its own, as what holds a stablecoin’s peg describes. Even the quote in a pair is a choice.

“Close” is a convention too. A market that trades around the clock has no closing bell, so the close of a candle is whichever print precedes a boundary that the data source picks. CoinGecko’s OHLC documentation counts its days in UTC, stamps each candle with its close time, and on the automatic setting picks the candle length from the range requested: 30 minutes for one to two days, four hours for three to thirty days, four days beyond that. By multiplication, a 14-period RSI on candles of those three lengths spans seven hours, 56 hours or 56 days. The formula, the asset and the data source are the same, and the question being answered is not.

None of this makes a venue’s own chart wrong. It makes it a description of that venue’s series: its participants, its quote asset, its volume. The same formula over another venue’s series returns another number, correctly computed from a different input.

Every construction above is built from trade prices and quantities: closes, highs, lows and volumes. None has a term for the bid, the ask or the gap between them. A line crossing on a chart therefore says nothing about what acting on it costs, which is the spread plus fees, and the fees piece sets out how, on a thin pair, the spread can exceed every commission combined.

Where indicators are honestly useful

The limits above do not make an indicator useless. They confine it to what it is: a calculation over a defined series. Three uses fit that.

Describing a regime after the fact. Whether the 12-period average sits above the 26-period, or the close above a longer average, is a precise and checkable statement about the recent past: prices have lately been higher than they were earlier in the window. It is a description of an interval that has already happened, and it trails that interval by construction.

Sizing volatility. Average True Range, from Wilder’s same 1978 book, takes for each period the greatest of three quantities (the high less the low, the high less the previous close, the low less the previous close, the last two as absolute values), which is the true range. It averages the true range over 14 periods, smoothed as RSI is. StockCharts’ account of it is explicit that it measures volatility and says nothing about direction. A rule that sets position size or stop distance as a multiple of recent ranges needs no view on which way price moves, which is why an indicator can serve it. It does rest on an assumption: that the coming periods will resemble the recent ones in range. Whether they do is the empirical question this section ends on. The indicator’s inputs are highs and lows, and on a venue’s own chart each is a single print, so one stray print can set a period’s whole true range.

Rules fixed in advance. A threshold chosen before a trade can be written down, applied identically each time and checked afterwards, so an indicator can be the unambiguous part of a plan: given the same series and the same parameters, two people compute the same number. Whether a particular rule suits a particular person depends on costs, risk limits and what they would do when it triggers, and none of that is in the indicator. No rule’s results in past data are offered here; they would describe the past like any other reading.

Of the three uses, none requires a view on direction; whether past inputs say anything about what comes next, in direction or in range, is an empirical question that the arithmetic cannot answer.

A reading order for any chart

Four questions come before the indicator, and each changes what the number can mean.

1. Which venue? Whose record is this: one venue’s own trades, or a data site’s blend of many? If a blend, under what published method, and does that method weight by volume?
2. Which pair? What is quoted against what: dollars, a token that tracks the dollar, euros? The same asset on the same venue can have several pairs, and each is its own series.
3. Which interval? How long is a candle, where does the day end, and so how much time do 14 periods cover? A setting counted in candles changes meaning when the candle length does.
4. Whose volume? The venue’s own, a site’s total of venues’ figures, or an adjusted figure? And does the indicator use volume at all? If it does, it inherits whatever is wrong with that figure.

With those answered, the indicator can be read for what it is: a summary of a particular series. If a chart cannot say which venue, which pair and which interval it draws on, the number on it was computed from something the reader cannot name.

None of this is an argument against computing an average. It is an argument for knowing what was averaged.