Trend following

Level 6 — strategy families. Assumes market types (level 3) and risk and sizing (level 5).

The claim

What has been going up tends to keep going up for long enough to pay for the times it doesn't.

Not "prices go up". The claim is about persistence: that a move already underway is more likely to continue than to reverse, by a margin wide enough to cover the losses from every move that reversed immediately.

That last clause is the whole strategy. A trend follower is usually wrong.

Why anyone believes it

Three mechanisms, none of them mystical, and they matter because a claim with a mechanism can be argued about while a claim with only a backtest cannot.

Information spreads slowly. A fact that changes what a company is worth does not reach every holder at once. Large holders cannot move quickly without moving the price against themselves, so they move over days or weeks, and that leaves a drift in one direction rather than a jump.

Someone has to be paid to take the other side. A person who wants out of a falling position needs a buyer. That buyer takes on risk and is compensated by price. This is a transfer, not an inefficiency, which is why it is harder to arbitrage away than a mispricing.

People buy what is going up. Money follows returns, and inflows buy more of what already rose. This one is the least respectable and the most reliably present.

Say the cost

Every one of those mechanisms is weaker now than it was in 1990. Faster information, cheaper execution and a crowded trade all cut the same margin. Trend following is the best-documented edge in public markets and one whose returns have measurably thinned. Both halves are true.

What it looks like

On a chart: a series of higher highs and higher lows, with pullbacks that stop short of the previous low. The eye is good at seeing this and bad at seeing how often it appeared and then failed — which is the reason this page ends in a study rather than a picture.

As logic, a trend follower is a channel rule. momentum_breakout is the one Arvo ships:

flowchart TD
    B["New bar closes"] --> Q{"Highest close<br/>of the last<br/>entry_period bars?"}
    Q -->|yes| G["Ask the risk gate<br/><small>sized, reduced or refused with a reason</small>"]
    Q -->|no| H{"Holding?"}
    G --> H
    H -->|yes| X{"Lowest close<br/>of the last<br/>exit_period bars?"}
    X -->|yes| S["Exit — no gate,<br/>nothing blocks shedding risk"]
    X -->|no| W["Hold"]
    H -->|no| W

Two numbers decide everything: how much strength you need to get in (entry_period) and how much weakness you accept before getting out (exit_period).

Arvo refuses one shape of this rule outright

If you set exit_period above entry_period, the engine rejects it:

exit period 60 must not exceed entry period 40; a rule that needs more evidence to leave than to enter gives most of a trend back before it admits the trend ended

A rule that is quick to buy and slow to sell is the most common beginner shape, because it feels patient. It is not patient, it is asymmetric in the wrong direction.

When it fails

It needs a trending regime. Arvo labels every trade trending up, trending down or ranging, so you can ask the question directly rather than guess: did this only work in one regime? A trend follower in a ranging market buys every false start and sells every false end, and pays the spread each time.

Most trades lose. A working trend follower wins maybe a third of its trades. The wins are much larger than the losses — which means the result rests on a handful of trades, and a handful is not a sample. This is why the verdict you get will often be Inconclusive rather than NotSupported: Arvo is telling you there is not enough evidence to separate skill from luck, which is a different statement from "it doesn't work".

The exit gives back the end of every trend. By construction, you leave after the reversal is visible. There is no version of this rule that exits at the top, and a backtest that appears to is a bug or a lookahead.

Costs bite twice. Trend rules trade often and enter on strength, which is when spreads widen. Arvo re-runs every finding with costs raised — half again the commission, twice the slippage — and ranks by what survived conservative costs first. Trend rules move down that leaderboard more than most.

Test it

Write rulesets/channel.json in your project:

{
  "name": "channel",
  "label": "Channel breakout",
  "premise": "A close at a 40 to 60 day high continues for long enough to pay for the false starts.",
  "interval": { "step": 1, "unit": "day" },
  "kind": {
    "kind": "grid",
    "rule": "momentum_breakout",
    "fixed": { "trade_size": 10.0 },
    "axes": { "entry_period": [40.0, 60.0], "exit_period": [10.0, 20.0] }
  }
}

Then Research view → pick channel → pick an instrument with bars → Run. Full steps in run a study.

Four configurations, so the winner is judged as the best of four, not as a single guess. That is deflation, and it is why the grid is small: every value you add raises the bar the winner has to clear.

Compare it against the control — sma_cross is in Arvo as a rule nobody disputes, not as a good idea:

import arvo

engine = arvo.connect()
found = engine.run_study("AAPL.RH", "sma_cross", author="script:trend-lesson")
print(found.verdict, found.read_this_first)

What to read, and in what order

1. The verdict, then the advice — before any number. If an advice item is blocking, the numbers below it are not a result yet.
2. The trade count. Under about thirty round trips, stop reading. A trend rule on two years of daily bars often lands here, and the fix is a longer window, not a looser entry.
3. The regime split on the Trades tab. If every winning trade is labelled trending up, you have learned something real and narrow: you have a rule that needs a regime, and you have no evidence about when that regime returns.
4. Against the benchmark. Holding the instrument is the bar. Check the dividend gap beside it — on a two-year test of a dividend payer, the gap can be most of your apparent margin.

Expect Inconclusive

On one instrument over a couple of years this rule will usually not produce enough trades to say anything. That is the lesson, not a failure of the test. The honest routes to a real answer are a longer window, or a panel — the same configuration over many instruments, which pools evidence instead of lowering the bar.

Watch

• (to be chosen) — what a trend looks like as it forms, in motion rather than in hindsight.
• (to be chosen) — why a low win rate can still be profitable, worked through with numbers.

Next

Mean reversion — the opposite claim, which is also true, in different conditions. Then why a backtest lies (level 7), which is where the two get separated from wishful thinking.