Trend, support and resistance

The vocabulary everybody uses. This chapter is about stating it precisely enough to see what it claims.

Trend

Loosely: a direction price has been moving. Precisely, it is usually defined as a sequence of higher highs and higher lows (an uptrend), or lower highs and lower lows (a downtrend).

That is a definition with real content — it is checkable — and it has two properties worth naming.

It is a description of the past, stated in the present tense. "The stock is in an uptrend" describes completed price action. It carries an implication about continuation, and that implication is doing all the work while remaining unstated.

It depends entirely on chapter 2's free parameter. Higher highs over what interval? The same data is an uptrend weekly and a downtrend daily, as chapter 2's worked problem showed. A trend is a property of the data plus your window, so a trend claim without a horizon attached is not yet a claim.

The honest version of the idea is this: prices have, in some markets and periods, shown autocorrelation — a tendency for moves to continue rather than reverse. That is an empirical claim, it has been tested, and chapter 9 reports what the testing found. It is a much narrower claim than "the trend is your friend".

Support and resistance

Support is a price level at which buying has previously been sufficient to stop a decline. Resistance is the mirror — a level at which selling has previously stopped a rise.

The idea is that these levels persist: that a price that stopped falling at 240 before will tend to stop falling at 240 again.

Why it might be true

Three mechanisms are proposed, and they are not equally good.

Memory and reference points. Holders who bought at 240 and watched it fall are, at 240 again, back to even — and the Loss aversion and the reference point chapter in Behavioural finance explains precisely why that matters to them. Kahneman and Tversky's observation that a person who has not made peace with his losses behaves differently is a real mechanism, and it predicts selling pressure at the level where people break even. This is the strongest of the three, because it rests on a documented behavioural regularity rather than on folklore.

Resting orders. If orders genuinely cluster at round numbers and previous extremes, there is more to transact against there. This is checkable in principle from order book data.

Self-fulfilment. If enough participants watch the same level and act on it, their action creates the effect. Note what this concedes: the level matters because people believe it does, not because of anything about the asset. That makes it real and fragile at once — it holds while the belief is widespread and fails without warning when it is not.

Why it might not be

The levels are drawn after the fact. You identify support by looking at where price stopped falling. That is not a prediction; it is a description, relabelled.

There are a great many candidate levels. A year of daily data has hundreds of local highs and lows, plus round numbers, plus previous closes. Some will coincide with later turning points by chance.

The failures are not counted. Levels that were breached are not remembered as support that failed; they stop being called support. A concept that only has examples of success has no error rate, and a claim with no error rate cannot be evaluated.

The methodological problem underneath

All of this is a specific case of what chapter 9 treats generally. The research literature defines it:

Data snooping occurs when a given set of data is used more than once for purposes of inference or model selection. If such data snooping occurs, any successful results may be spurious because they could be obtained just by chance with exaggerated significance levels.

Drawing a support line is searching one dataset for a level that worked, then presenting the level as a finding. The same paper notes a subtler version: a dataset can be repeatedly used to search over families of rules, markets, estimation periods and model assumptions — and that collective data snooping "is potentially even more dangerous because it is not easily recognized by each individual researcher."

Which applies to a million chart-watchers examining the same price history as much as it does to academics.

Working the problem

"Strong support — price bounced off it three times."

The case that it is informative. Three touches is more than one, and the reference-point mechanism gives a reason to expect behaviour at a remembered level. If large holders bought near that price, they have a documented tendency to act when returned to break-even. If orders cluster there, there is genuinely more liquidity to absorb selling. And if many participants are watching the same level, their collective behaviour can produce the effect regardless of whether it was ever "real".

The case that it is not. Begin with how many levels exist. A year of daily data contains roughly 250 bars, and therefore scores of local lows, plus every round number in the range, plus notable previous closes — easily a hundred candidate levels. In a random series, some of those will be touched three times without bouncing being meaningful, simply because price oscillates and levels are dense.

And the level was identified because it had three touches. You searched the history for a level with that property and found one, which is guaranteed to succeed and tells you nothing. That is data snooping in its most literal form.

What evidence would settle it. Three things, and all are obtainable in principle:

1. A base rate. Across many instruments and years, define support mechanically in advance, then measure what fraction of such levels held on the next touch. Compare with what a random series produces. Without that comparison there is no finding.
2. Prediction before the fact, on data not used to find the level — the replication approach chapter 9 describes.
3. Order book evidence that volume actually rests at those levels, which would distinguish the mechanism from the story.

In the absence of all three, "it bounced three times" is an observation about the past presented as a property of the future. It may still be right. It is not yet evidence.

The point

A trend is usually defined as higher highs and higher lows, which is checkable but is a description of the past carrying an unstated implication about continuation — and it depends on a window you chose, so a trend claim without a horizon is incomplete. Support and resistance have one strong proposed mechanism in the reference-point behaviour documented in behavioural finance, one checkable one in resting orders, and one self-fulfilling one that concedes the level matters only because people believe it. Against that, levels are drawn after the fact, candidates are dense enough that chance produces apparent bounces, and failures are quietly reclassified rather than counted.