06 · Technical Analysis: Critique and Validation
The previous articles taught you how to "use" technical analysis; this one pours cold water on it: does technical analysis actually work? What does academia say? Why do some people make money and others lose on the same pattern? And — when is a signal "genuinely effective" rather than "looks effective"?
💡 Master Principle
Remember one master principle first: technical analysis's standing in academia is "some real effects exist, but far from enough to support most folk usage". This article does not deny the usefulness of technical analysis, but demands a verifiable audit of every claimed "power".
1. Supporting Evidence: The Parts of Technical Analysis That Are "Real"
Critics often dismiss technical analysis wholesale, but decades of finance research have actually found a few effects "repeatedly confirmed" — and they happen to be the theoretical footing for certain technical-analysis usages.
1.1 The Momentum Effect
- Finding: Jegadeesh & Titman (1993) and a large follow-up literature confirmed: stocks that outperformed over the past 3–12 months tend to keep outperforming over the next 3–12 months; and vice versa. This is one of the most robustly replicated anomalies in finance.
- Link to technical analysis: the momentum effect is the most important scientific basis for "trend following" (bullish MA alignment, breakout buying, trend patterns) — trends genuinely have inertia; going with the trend is not statistically stupid.
- Boundary: momentum reverses over short horizons (days) and very long horizons (years) (short-term and long-term reversal effects), and momentum strategies suffer deep drawdowns during momentum crashes (sharp market turns) — "following the trend" is not unconditional profit.
1.2 Anchoring and Self-Fulfillment of Support/Resistance
- Anchoring effect: behavioral finance confirms that humans naturally anchor to price levels like "round numbers, prior highs/lows, historical averages", and traders use them as decision references — giving support/resistance a real behavioral foundation;
- Self-fulfillment: when enough people believe "20 is support", they place bids around 20, objectively creating the support — part of why support/resistance "works" is that believers turn it into fact with real money;
- Research corroboration: Osler (2000) studied the distribution of take-profit/stop-loss orders in forex and found orders do cluster near "round numbers and prior highs/lows on the chart" — support/resistance does have real order-flow footing, but that also means it gets run through systematically (see Section 3).
1.3 Established Findings on Volume Anomalies
- Expansion accompanies large moves: academic research broadly supports volume-price common sense such as "volume correlates positively with absolute returns" and "expanding breakouts are more real than contracting ones";
- Volume-price divergence carries information: price making new highs on shrinking volume is read as "insufficient participation", with statistically higher odds of subsequent underperformance;
- Limitation: volume research is "descriptive" — it shows "volatility is greater when volume expands", but once you turn "expansion" into a tradeable entry signal, the after-cost edge is often negligible.
2. Critical Evidence: Why Technical Analysis Is "Not That Magical"
2.1 Random Walk and the Weak-Form Efficient Market Hypothesis
- Random walk: price changes cannot be predicted from historical prices (next direction ≈ a coin flip) — the fundamental challenge to technical analysis; if historical patterns cannot predict the future, technical analysis carries no information;
- Weak-form EMH: proposed by Eugene Fama — if a market is weak-form efficient, then all historical price and volume information is already reflected in the current price, and any strategy based only on historical price/volume cannot earn excess returns;
- Realistic conclusion: EMH is not "completely right" — markets contain many anomalies (momentum, low volatility, etc.), but whether those anomalies survive transaction costs and real capital capacity is another matter — research leans toward "efficient but imperfect".
2.2 The Difficulty of Beating a Benchmark with Technical Strategies
| Strategy | Typical academic backtest result | Conclusion |
|---|---|---|
| Dual MA (golden/death cross) | Long-run returns roughly equal to or worse than buy-and-hold, with high costs | The "effectiveness" comes from bull-market exposure, not timing |
| MA timing (e.g., MA200 filter) | Positive contribution in some markets/periods, but drawdowns and costs eat most of the edge | Fragile, parameter-sensitive |
| Classic patterns (head and shoulders, double top) | Some statistical significance (Brock et al. 1992 supported it), but later studies show severe out-of-sample decay | Strong period-of-publication effect, hard to replicate |
| Oscillators (RSI/KDJ overbought/oversold) | Decent in ranges, repeatedly slapped in trends | Regime-dependent, and the regime itself is hard to predict |
💡 The Academic Consensus in One Sentence
The academic consensus in one sentence: technical-analysis strategies in backtests are "occasionally effective, generally fragile, cost-sensitive, and rarely beat buy-and-hold consistently".
2.3 Landmark Studies (The Tug of War)
| Study | Finding | Meaning for technical analysis |
|---|---|---|
| Fama (1970), weak-form EMH | Historical price information yields no excess returns | The "theoretical basis" of technical analysis shaken |
| Brock, Lakonishok & LeBaron (1992) | MA rules showed statistically significant predictive power on the Dow | Technical analysis gained "academic legitimacy"; cited endlessly |
| Later out-of-sample tests (Sullivan et al. 1999; Hsu & Kuan 2005) | After data snooping and out-of-sample testing, the edge decays sharply | Early "effective" results shown to be partly cherry-picked |
| Jegadeesh & Titman (1993), momentum | Intermediate-term momentum is real and robustly replicated | Gave "trend following" a solid behavioral foundation |
| Osler (2000), order distribution | Stop-loss/take-profit orders cluster at round numbers and pattern levels on charts | Support/resistance has a real order-flow foundation |
| Lo (2004), Adaptive Markets Hypothesis | Market efficiency is dynamic: strategies cycle effective → crowded → dead | Explains why signals "sometimes work, sometimes don't" |
The overall picture from these studies: technical analysis is not "pure superstition" (momentum, support/resistance, and volume-price all have real behavioral/microstructure foundations), but it is nowhere near the folk-marketed "holy grail" — its true effects are "weak, fragile, and dynamically varying with crowding".
2.4 Survivorship Bias and Publication Bias
- Publication bias: academia publishes only the "winners" — "the wins get published; the losses get written by nobody". A researcher can test 100 rules; the 5 that happened to win become a paper, the other 95 go into the drawer (the "file drawer problem"). Readers see a filtered, optimistic subset;
- Survivorship bias: backtests use only stocks/instruments alive today; delisted and zeroed-out ones are excluded — so the "historical success rates" of technical analysis (especially on small caps and junk stocks) are systematically overstated;
- Data snooping: retest parameters repeatedly on the same history and you will always find "perfect parameters" — perfect only for that history (overfit), with zero predictive power for the future. For every "80% win-rate strategy" you see, first ask: was it tested, or was it selected?
How an "amazing backtest curve" is born:
100 raw rules → backtest → discard 95 → publish the 5 that won
↓
Readers conclude "technical analysis beats the market on average"
The true distribution may be "the luckiest 5%"3. The "Self-Fulfillment" Problem: More Users, More Effective or More Broken?
3.1 The Two Faces of Self-Fulfillment
- Positive self-fulfillment: when moving averages, support/resistance, and classic patterns are watched by huge numbers of traders, "bids clustering at support" emerges as behavior and the signal gets "more accurate" short term — which is why widely known signals (the 60-day MA, round numbers) show up more clearly in retail-dominated markets (e.g., A-shares);
- Negative self-fulfillment (crowded trades): when a signal is used by too many, a "front-running effect" appears: everyone buys at the same level, price instantly jumps past it, and latecomers get worse fills; worse still is exploitation by operators — deliberately faking moves just above the support everyone waits on, or sweeping stops just below it before pumping ("stop hunting"), turning the "effective signal" into a trap that harvests retail traders.
3.2 Typical Signatures of Crowded Trades
| Signature | Description |
|---|---|
| Overly popular signal | Golden crosses everyone knows, trendlines everyone draws — front-running and stop-hunting both intensify |
| Overly tidy levels | When stops pile up near a support, the odds of the market "hunting stops specifically" rise markedly |
| Overly synchronized timing | Around earnings/delivery dates/time windows, consensus expectations create "sell the news" reversals |
💡 Conclusion
Conclusion: self-fulfillment makes a signal "more effective short term, more crowded long term"; the two alternate, turning signal effectiveness itself into a curve of decay and reformation — you can never assume a signal stays effective forever.
4. How to Scientifically Validate a Technical Indicator (Actionable Steps)
No academic background needed — with a market terminal and a spreadsheet you can give any indicator a "scientific health check". The goal is not to prove it 100% effective, but to measure its true win rate, payoff ratio, and drawdown, and judge whether it is worth using.
4.1 The Seven-Step Validation Method
① Write down the complete rule definition (reproducible by a program or anyone)
② In-sample backtest (2015–2021, parameter A)
③ Out-of-sample backtest (2022–present, same parameter A)
④ Parameter sensitivity test (parameter A ±20%, ±50%)
⑤ Rerun with transaction costs and **<mark>slippage</mark>** included
⑥ Benchmark against "buy-and-hold" and "random signals"
⑦ Conclusion: still effective out of sample? robust to parameter changes?4.2 The Key Question at Each Step
| Step | Question to answer | Common pitfall |
|---|---|---|
| Define the rule | What exactly is a "golden cross"? (close or intraday price? confirmed after how many bars?) | Vague rules → backtest not reproducible |
| In-sample | How are returns under this parameter? | Testing only one bull run inflates results |
| Out-of-sample | On history never used for tuning, do results hold? | In-sample-only effectiveness = overfit; discard |
| Parameter sensitivity | With 20 changed to 10 or 40, does it still profit? | Only parameter 20 profiting = fitted to noise |
| After costs | Does it still profit after 0.1%–0.2% fees + slippage per trade? | High-frequency small-signal strategies get eaten alive by costs |
| Benchmark comparison | Better than "buy-and-hold" and "random entries"? | Beating the benchmark counts; beating zero does not |
| Overall judgment | Out-of-sample return + parameter robustness + positive after costs — all three must pass | Fail any one → treat as "ineffective" |
4.3 Three Veto Conditions
- Out-of-sample failure — the strategy works only on the history used for tuning;
- Parameter sensitivity — a strategy that flips from profit to loss on ±10% parameter shifts is no strategy;
- Negative after costs — high gross returns that turn to losses after costs and slippage: the most common reality of retail strategies.
4.4 A Complete Validation Example (Dual-MA Golden Cross)
Take "buy when MA20 crosses above MA60, sell when it crosses below" through the seven steps:
| Step | Execution | Typical result and reading |
|---|---|---|
| Definition | Golden/death cross confirmed on close, executed at next open | Reproducible rule; passes |
| In-sample | CSI 300 index, 2015–2021 | 9% annualized vs 6% buy-and-hold — looks good |
| Out-of-sample | 2022–2025, same rule, same parameters | 4% annualized vs 5% buy-and-hold — clear decay |
| Parameter sensitivity | MA pairs changed to 10/40, 30/90, 50/150 | Only 20/60 profits, others much worse — overfit signature |
| After costs | 0.2% per turn (commission + slippage), ~15 turns/year | Another 1.5 points off annualized; out-of-sample edge gone |
| Benchmark | vs random timing (coin-flip entries/exits) | No significant difference |
| Conclusion | Out-of-sample failure + parameter sensitivity + no after-cost edge | Verdict: ineffective; discard |
The point of this example: the folk claim that "golden-cross strategies are profitable" almost never survives this process. Not because the process is too harsh, but because the real costs and real noise of trading markets simply cannot accommodate most "beautiful-looking" signals.
As an aside: if you can't write backtests, use the dumbest effective method — manual journaling: from today on, log every signal's trigger time, price, and reason in a spreadsheet; after a month, check the "5-day/10-day returns after each trigger". After a few hundred samples you will know the signal's "true win rate" far better than by feel.
5. The Reasonable Position of Technical Analysis
5.1 An "Explanation Tool", Not a "Prediction Tool"
- What it can do: describe the market's current regime (trend/range/exhaustion), what the bull-bear structure looks like, which price levels have real money attention — background information that lowers decision difficulty;
- What it cannot do: tell you whether price rises or falls tomorrow, exact buy/sell points, or price targets. Any technical analysis claiming to do these exceeds its capability boundary;
- The right posture: treat technical analysis as a "GPS map" (where am I, what terrain lies ahead), not an "oracle" (what will happen tomorrow).
5.2 As an Entry-Timing Filter for a Trading System
Technical analysis's real role in professional trading is usually a "filter", not the "engine":
Complete trading system = Direction (fundamentals/macro trend judgment)
+ Trigger (technical entry signal) ← only decides "when"
+ Risk control (position sizing, stops, payoff ratio)
+ Execution (discipline and consistency)| Role | Technical analysis's duty | Duty it should NOT carry |
|---|---|---|
| What to trade | Not its job (fundamentals/macro set direction) | Don't pick instruments by "pretty patterns" |
| When to enter | Breakout/retest/exhaustion signals give the trigger | No guarantee of profit after the trigger |
| When to exit | Stop levels, trend-break signals | Don't hold and hope on "theoretical price targets" |
| How big a position | Not its job (set by ATR and risk budget) | Don't size up because "the signal is strong" |
💡 Positioning in One Sentence
Positioning in one sentence: technical analysis's job is to "find a good entry timing in the right direction", and it always defers to position management and stop-losses.
6. Why Many People Lose Money with Technical Analysis
6.1 Mistaking Correlation for Causation
- Seeing "price often rises after a golden cross", they conclude "the golden cross causes the rise" — in fact the golden cross is just the trend's echo (a lagging indicator); when trends rise, golden crosses naturally abound; taking the echo for the cause gets you into counter-trend trades at market turns;
- Likewise, "this pattern worked last time" does not imply "it will work this time" — tiny samples + memory bias (remembering only the wins) manufacture false confidence.
6.2 Ignoring Position Sizing and Execution
- Payoff ratio and position size are the variables that shape the equity curve: a 55%-win, 1:1 payoff system loses long-term to a 40%-win, 2:1 payoff system — technical analysis only touches the win-rate term;
- The typical retail loss path: the signal is fine, but position too heavy (−10% per loss) + stops not enforced (cutting only at −30%) + winners dumped too early (+3% and out) — together these three turn a "positive-expectancy signal" into a "negative-expectancy account";
- Technical analysis tells you nothing about "how much to bet", and that is precisely the part that decides survival.
6.3 Treating Indicators as Holy Grails
- The "grail mentality": believing a "90% win-rate indicator" means financial freedom — so they keep switching indicators and optimizing parameters, sinking ever deeper into the overfitting loop; behind every seemingly perfect strategy lies a historical curve that has already stopped working;
- Signal abuse: run 5 indicators and one always supports your idea — more indicators mean more "selective belief"; in the end you trade not a system but your own emotions;
- The only antidote: apply the Section 4 validation method to every strategy you use, and accept that "a good technical system is a thin edge + strict execution". Thin edge × long-term discipline = stability; an "invincible indicator" = guaranteed loss.
⚠️ Risk Warning
Technical analysis is a probabilistic tool, not a certainty tool: the real effects supported by research (trend effects, support/resistance anchoring) are often heavily diluted once transaction costs, slippage, and parameter overfitting are counted; most "beating-the-market" backtested strategies are contaminated by publication bias, survivorship bias, and data snooping, and decay out of sample; crowded trades on homogeneous signals can even harvest their own users. Any technical signal can only stay effective long-term when combined with position management, stop discipline, and consistent execution — treating technical analysis as a holy grail, sizing by signal strength, or ignoring costs are all classic paths to losses. This article is for educational purposes only and does not constitute investment advice.