Why technical indicators cannot prove 90% odds

Stacking indicators can make a trade look unusually convincing. RSI turns higher, MACD crosses, the price moves above a regression line and a moving average confirms the direction. When several signals point the same way, it is tempting to say the probability of success has become very high.

The problem is that three indicators are not automatically three independent pieces of evidence. They may be different calculations built from the same price history. Agreement can still be useful, but it does not justify attaching a number such as “90 per cent” without a properly defined and tested process.

More indicators do not automatically mean more independent evidence.

Three signals may still be one source of information

RSI measures the balance of recent gains and losses. MACD compares moving averages of the same price series. A regression line estimates the direction or relationship within that price series. Their formulas differ, but their inputs overlap heavily.

This matters because probabilities do not compound simply because several indicators agree. If three independent tests each added genuinely new information, combining them might improve confidence. If all three react to the same trend, however, the second and third signals may mostly repeat what the first one already detected.

Think of three weather apps that all use the same forecasting model. Seeing the same rain forecast three times does not give you three independent forecasts. Technical indicators can have the same problem: the chart looks more crowded while the underlying evidence has barely changed.

Inputs that come from somewhere else may add more information. Trading volume, company earnings, liquidity conditions and market structure are not automatically independent either, but they at least require a different question from three transformations of closing prices. The test is not how many lines appear on the chart. It is how much genuinely new information each line contributes.

Count independent transmission channels, not indicator labels

A chart can look heavily confirmed when RSI, moving averages, regression lines and volume all point in the same direction. The useful question is whether the signals measure independent information or repeatedly transform the same price series.

Separate the evidence into four layers:

• Price: trend, volatility, volume and market breadth
• Discount rate: long yields, real yields and inflation expectations
• Credit: corporate spreads, issuance, bank funding and lending standards
• Real economy: earnings revisions, employment, energy and freight costs, inventories and orders

Five indicators inside the price layer do not carry the same weight as deterioration across all four layers. The second pattern offers a clearer mechanism for a regime change. The broader market-regime framework shows how to test those transmission channels without turning the checklist into a mechanical forecast.

A probability needs a denominator

Before accepting a claim that a setup works 90 per cent of the time, ask what was actually measured.

Was it tested on every signal or only selected examples? On ASX shares, broad-market ETFs, US equities or one unusually strong chart? Was the entry taken at the closing price, the next session’s open or a later price that would not have been available in real time? How long was each position held? What counted as success? How many trades were included?

Without those details, “90 per cent” is not a probability that an investor can evaluate. It is a number without a denominator.

Win rate is also only one part of a strategy. A method that wins often can still lose money if its occasional losses are much larger than its typical gains. A more useful starting point is expectancy: the average gain from winning trades, minus the average loss from losing trades, adjusted for how often each occurs and for trading costs.

A live screener or paid signal service needs the same denominator. A handful of shares that rose after being mentioned, selected subscriber testimonials and a prominent risk disclaimer do not reveal the service’s expectancy. The comparison needs the complete signal record, the losses and the drawdowns as well as the winners.

A precise similarity score is a claim, not a cause

A statement that today’s chart is 97 per cent similar to a past crash cannot be evaluated without the method. Investors need the selected windows, scaling, variables, search universe, rejected matches and out-of-sample record.

Visual similarity does not establish identical causes. Policy settings, credit conditions, market structure, earnings and positioning can differ even when two price paths resemble each other. A date and index target should remain a scenario until funding, credit and liquidity data confirm the mechanism.

Why backtests often look better than live trading

Backtesting is valuable, but it is easy to produce a result that is cleaner than the experience an investor will have in the market. The most common problems are:

• Data snooping. If enough indicators, settings and entry rules are tested, something will eventually look impressive by chance.
• Overfitting. A rule can be tuned so precisely to past data that it captures historical noise rather than a repeatable relationship.
• Look-ahead bias. A test may accidentally use information that would not have been available when the trade decision was made.
• Survivorship bias. Testing only securities that still exist today can exclude failed, delisted or acquired companies and make the historical universe look stronger.
• The execution gap. Brokerage, bid–ask spreads, slippage, liquidity and tax treatment can change a paper result materially. A rule tested on a highly liquid ETF may not transfer unchanged to a thinly traded small-cap share.

The data-snooping problem is not theoretical. Sullivan, Timmermann and White’s study examined how apparent technical-rule performance changes when researchers account for the full universe of rules that was searched. CFA Institute’s Backtesting & Simulation reading likewise treats survivorship bias, look-ahead bias, structural breaks, fat tails and realistic implementation as central parts of a credible test.

Paid screeners need a complete record, not selected winners

A credible live record should retain every signal when it was generated, including failed and stopped-out trades, securities that were halted or too illiquid to enter, and every later rule change. It should use a price a subscriber could realistically have obtained and include brokerage, bid–ask spreads, slippage and relevant tax effects. The operator’s own fill is not automatically available to every subscriber.

Advertising warnings do not prove performance. They should be read alongside the headline return, maximum drawdown, losing streaks, benchmark and evidence that the same rules were applied before the outcome was known.

Licensing and advertising are part of due diligence

In Australia, the legal question depends on the conduct, not whether a service calls itself education or signals. ASIC says that a recommendation or opinion intended, or reasonably regarded as intended, to influence a decision about financial products may amount to financial product advice. A person carrying on a financial-services business generally needs an Australian financial services licence, authorisation or an applicable exemption.

That does not mean every market discussion needs a licence. Factual information can remain factual, while a recommendation, arranging a transaction or receiving payment to promote a product can change the position. Before paying for a service, check the provider’s identity and authorisation on ASIC’s professional registers , read its price, refund terms, conflicts and risk disclosures, and treat unusually precise figures or claims that returns are certain as statements requiring evidence rather than proof.

AI-generated candidate lists are research leads, not evidence

An AI or automated stock list can speed up the first pass, but it can also return a suspended security, an outdated article or a company whose connection to the advertised theme is only superficial. A candidate should not become a trade because it appeared in a generated list.

Verify that the security is currently tradeable, read the issuer’s latest ASX announcements and confirm the economic link. Look for a disclosed contract, revenue exposure or balance-sheet effect rather than a keyword match. Record the source date and reject the idea when the link cannot be established.

Leveraged inverse products can lose even when the directional view is broadly right

Most leveraged and inverse exchange-traded products target a multiple of one day’s return. Their multi-day result can diverge sharply from the simple multiple of the benchmark’s return over the same period.

Suppose an index falls 10 per cent on day one and rises 11.11 per cent on day two. It returns to its starting value. A daily -3x product that perfectly meets its objective would rise 30 per cent, then fall about 33.33 per cent. A starting value of 100 becomes 130, then roughly 86.7. The index is flat across the two days; the product has lost about 13.3 per cent.

Australian leveraged and inverse ETPs do not all use one identical borrowing structure. ASX notes that leveraged ETPs may use borrowing or over-the-counter derivatives, magnify gains and losses, increase volatility and carry product-specific liquidity, financing and counterparty risks. The product disclosure statement controls the details for each product.

A tight stop does not remove this risk. Gaps, liquidity, tracking error and execution can prevent an exit at the planned price. Position size, borrowed money, holding period and maximum account loss matter before the chart target.

What would weaken this view

The counter-signal is a technical rule that survives a process designed to expose false confidence. The view would weaken if the rule and thresholds were fixed before testing, the full set of attempted rules and failed cases were disclosed, and the result remained useful on unseen data and in other markets.

The result would also need to retain a positive expectancy after realistic brokerage, spreads and slippage, while keeping drawdowns and losing streaks within the stated risk limits. Those are observable conditions that would change the view. A selected chart or one successful period is not enough.

Seven questions to ask before believing a high-probability claim

A useful claim should survive questions such as:

1. Were the signal rules, trade timing and exit rules fixed before the results were examined?
2. Are success, failure, holding period and sample size defined clearly?
3. Does the test disclose the full set of securities, periods and rule variations that were tried?
4. Were the data used to develop the rule separated from the data used to validate it?
5. Were brokerage, spreads, slippage and other implementation costs included?
6. Does the report show expectancy, drawdown and losing streaks as well as win rate?
7. Does the result remain useful across different markets, volatility environments and trend conditions?

If several of these answers are missing, the percentage should be treated as an untested hypothesis or a marketing claim, not as a dependable estimate.

Additional questions for paid or automated systems

• Was the rule published before the signal occurred?
• Is there a timestamped log of every signal and every rule change?
• Were halted, illiquid and untradeable securities retained in the record?
• Are brokerage, spreads, slippage and relevant tax effects included?
• Are failed trades, drawdowns and losing periods shown with the winners?
• Could subscribers have obtained the same entries and exits as the operator?
• Are the provider’s authorisation, price, refund terms, conflicts and risk disclosures clear?
• Was each AI-generated candidate checked against current issuer announcements and trading status?

What technical indicators are actually good for

None of this makes indicators useless. Their real value is more practical than predictive.

Indicators can turn a vague impression into a repeatable rule. They can help an investor define when a trend is strengthening, when momentum is fading, where a position should be reduced and how much risk to take. Used this way, an indicator supports decision discipline rather than pretending to remove uncertainty.

The distinction matters. “RSI, MACD and the regression line agree, therefore this trade has a 90 per cent chance of success” is an unsupported probability claim. “When these conditions occur, I will enter at a defined time, risk a defined amount and exit under a defined rule” is a process that can be tested.

The practical guide to using MACD for trend confirmation and risk control turns that distinction into caution, partial reduction, no-trade and re-entry conditions.

A useful risk-control rule does not prove the forecast has an edge

Two different questions are often bundled into a technical strategy. The first is whether the signal predicts market direction with repeatable positive expectancy. The second is whether the procedure limits damage when the signal is wrong. Position size, a maximum account loss, pre-defined invalidation and permission not to trade can improve the second without proving the first.

Even an attractive win rate cannot protect an account that repeatedly risks most of its capital. Limiting each loss to a small, pre-defined part of the account preserves the ability to take the next observation and reduces the chance that one gap or false break ends the process. The account-level loss limit should come before selecting a percentage stop for one security.

Invalidation should also be written before entry. A failed breakout, a support break, missing confirmation from volume or breadth, or a time limit can all be observable conditions. An alternative scenario, a re-entry rule and a no-trade outcome are needed as well. Without them, a trader can redraw the chart after each move and call the changed explanation the original plan.

A stop is not an automatic guarantee. Opening gaps, thin order books, halts, daily resets in leveraged products and tracking error can move the execution away from the planned level. A rule that uses a break in US index support to reduce an Australian semiconductor or leveraged-ETP position is testable, but it needs aligned timestamps, the futures-versus-cash distinction, currency effects, whipsaw frequency and re-entry costs.

The clean evaluation is therefore two-stage. First ask whether the sizing, invalidation and waiting rules reduce ruin risk in practice. Then inspect a complete chronological signal record for after-cost expectancy, maximum drawdown, losing streaks and re-entry costs. Passing the first test means the method contains useful survival controls. It does not mean its predictive edge has been established.

Market regimes change what a signal means

A technical rule does not operate in a vacuum. Trend-following signals can behave differently in a persistent trend than in a choppy, range-bound market. A signal tested during calm conditions may perform differently when volatility rises, liquidity thins or correlations move sharply higher.

That is why a single backtest average can conceal more than it reveals. Investors should examine how the rule behaved across distinct environments, not only across one long period. The aim is to understand when the process has historically been useful, when it has struggled and how much loss can occur before that becomes obvious.

This is the same discipline used in a broader market-regime framework : define the environment, identify the mechanism and state what evidence would invalidate the working view.

Conditions that would strengthen or weaken a crash thesis

A severe-downturn thesis becomes more credible when several mechanisms deteriorate together:

• corporate spreads widen and issuance becomes difficult;
• bank funding costs rise and lending standards tighten;
• weak breadth persists for weeks rather than a session;
• earnings downgrades spread across sectors;
• arrears, refinancing stress and forced sales rise; and
• energy or freight costs pass into margins and inflation.

The thesis weakens when credit remains available, funding markets function, earnings revisions stabilise and breadth improves. A steepening yield curve or a familiar chart shape is not sufficient on its own.

A more credible testing standard

A sensible process does not need to be complicated, but it should be difficult to fool.

Start with a clear hypothesis. Write the entry, exit and position-sizing rules before looking at the final result. Separate the period used to develop the idea from the period used to test it. Include realistic costs. Report the distribution of outcomes, not only the average or win rate. Test other securities and different market conditions. Finally, observe the rule prospectively, on paper or at very small size, before relying on it.

The goal is not to prove that a chart pattern is certain. It is to find out whether the process has a positive expectancy, whether the evidence is robust enough to matter and whether the drawdowns are tolerable for the person using it.

When several indicators line up, begin by counting independent inputs rather than signals. Then ask whether the rules were fixed in advance, whether the result survived unseen data and whether costs and changing market regimes were included. That will not produce the certainty promised by a “90 per cent setup”, but it gives an investor something more useful: a repeatable decision process with limits that can be measured.

Official investor guidance

• SEC investor bulletin on leveraged and inverse ETFs
• FINRA overview of leveraged and inverse exchange-traded products
• ASIC, Discussing financial products and services online
• ASIC, RG 234 Advertising financial products and services
• ASX, Investing in exchange-traded products