Someone sends you a chart of Harbor Coffee, our fictional coffee company. Three rises reach $68, $72 and $68. The message says, “Head and shoulders. A fall is coming.”
Cover the chart's right-hand edge. One reader sees a reversal taking shape; another sees a pause before gains. The prices are the same. The stories differ.
Before copying the forecast into your research note, separate the shape from the prediction. You can agree about what happened without agreeing about what comes next.
A shape is a proposed rule
A multi-bar chart pattern is a shape formed by a sequence of prices. You read across many bars rather than interpreting one candlestick.
A continuation pattern suggests that a trend will resume after a pause. A reversal pattern suggests a change of direction. The names describe the forecast; they do not settle whether it is right.
Head and shoulders and double tops suggest a reversal after a rise. Flags suggest continuation after a sharp rise or fall; cups and handles suggest continuation of an uptrend.
| Pattern | Shape | Completion |
|---|---|---|
| Head and shoulders | Three peaks; middle highest | Below neckline |
| Double top | Two similar peaks | Below middle low |
| Flag | Short, parallel-sided pause | Break with prior trend |
| Cup and handle | Rounded bowl, small pullback | Above handle resistance |
In head and shoulders, the neckline joins the two lows between the peaks. A break below it completes the pattern. The other patterns also need a break of support or resistance, areas where earlier price moves paused or reversed.
“Looks like head and shoulders” is a description. A test needs a rule that two readers can apply the same way, including what counts as a peak and a break.
Freeze the chart before the ending
Suppose Harbor has these six daily closes after year 3: $64, $68, $65, $72, $65 and $68.
The first $68 suggests a left shoulder and $72 the head. The latest $68 could become the right shoulder, but it is still the chart's last point. You have not yet seen it turn down. The two intervening lows are both $65, so the neckline is horizontal.
Before seeing another price, set this example's rule: the next close below $65 completes the break. Call the forecast correct if the price closes lower still two sessions after that break. Without a deadline, “it will fall” can wait forever to be right.
The next close is $64. The rule has triggered. From there, imagine two endings: A falls to $62, then $60; B recovers to $67, then $70. Both paths share every price through the break.
Only A passes the two-session test. These are two possibilities, not a 50% success rate. A real test must keep recoveries like B among its failures.
Keep the original shoulders fixed, just as you kept the support zone fixed earlier. A later fall cannot make the earlier forecast obvious.
Ask what studies actually found
In their 2000 study, Lo, Mamaysky and Wang used computer rules to identify patterns in US stocks from 1962–1996. For several patterns, the mix of daily returns that followed differed from the mix overall. The patterns contained some information about what came next.
That is weaker than showing a profitable strategy. You still need entry and exit rules you could execute, with gains large enough to cover costs. A statistical difference does not pay the trading bill.
A 2012 study by Bajgrowicz and Scaillet examined daily Dow Jones Industrial Average closing prices from 1897–2011. They tested a different family of rules, including moving-average rules.
Knowing which rules had won in the past did not reliably identify the next winners. Even low trading costs erased the apparent historical advantage in their tests.
The studies ask different questions. Together, they leave room for measurable information in prices while showing how much more evidence a trading claim needs.
Read the last sentence of the linked Lo abstract. Does it establish information in their sample or profits from the chart in your message? Only the first claim follows from their finding.
Make a lucky result work harder
A backtest asks how a strategy would have performed on historical data. Data snooping, also called data mining here, means searching many rule variations on the same history, then presenting the winner as though it were the only test.
With a fair coin and independent flips, a ten-flip trial chosen in advance has a 1 in 1,024 chance of producing ten heads.
Run 1,024 independent trials and you expect 1,024 × 1/1,024 = 1 all-heads result on average. One batch could produce none or several. Chart rules share price inputs, so they are not independent coin trials, but the search problem remains.
If someone shows you only the winning sequence, the missing number is how many tries it took.
Trying dozens of neckline thresholds, holding periods or stock lists gives chance more opportunities to look clever. The discarded rules belong beside the winner when judging the evidence.
Make the test match real decisions
Three checks keep the test tied to what you could have known:
- Keep the clock honest. Look-ahead bias means using information before it was available. Harbor's final $68 is only a possible shoulder at session 6; treating it as a confirmed peak imports a later price.
- Keep the missing companies. Survivorship bias in backtests means testing only stocks that survived. Include failed and delisted companies that met the rule at the time.
- Leave some history untouched. An out-of-sample test uses observations never used to choose or tune the rule. If you change the rule after seeing those results, that history becomes practice data. The revised rule needs a fresh test.
Timing applies to trades too. Harbor's $64 close tells you the rule fired; it does not promise you could then enter at $64. A backtest needs an execution price available after the signal is known.
Say a separate $1,000 test position gains $6 before $8 of trading costs. The net result is $6 − $8 = −$2. Those round-trip costs cover entry and exit: fees, the bid–ask spread and worse execution prices than the test assumed.
The price call can be right while the trade loses money. Even before costs, counting wins misses their size: a few large losses can outweigh many small gains.
The earlier relative-strength comparison showed why a stock's result needs context. Apply that principle to a rule: a predicted fall has to add something beyond how often comparable stocks fell over the same horizon without the pattern. Otherwise, it may just be describing a falling market.
Return to the screenshot. The sender reveals that Harbor later fell, but supplies no rule set in advance, failed cases or after-cost results. Which claim can your research note accept?
“The prices resemble head and shoulders” can stand. “This predicts a profitable trade” still needs evidence. Keep the description and leave the forecast out. That distinction carries into using technicals alongside fundamentals.
In short
- A named shape becomes testable when its rules and deadline are fixed before the outcome.
- A convincing example is not a representative sample.
- A test needs unseen data, failed cases, discarded rules and realistic costs.
- Finding information in prices is a lower bar than finding a profitable trade you can execute.