Mean reversion¶

Level 6 — strategy families.

The claim¶

Price pushed unusually far from a fair value comes back to it.

The direct opposite of trend following, and also true — in different conditions. Where a trend follower buys strength, a mean reverter buys weakness, on the argument that the weakness was an overshoot rather than information.

Everything depends on what "fair value" means, and the honest answer is that it is always a proxy. Arvo's intraday rule uses VWAP — the volume-weighted average price since the session opened, the average price at which the day's business actually happened. It is a reasonable proxy for "what this thing traded at today" and it is not a valuation.

Why anyone believes it¶

Someone needed to sell, and need is not information. A fund meeting redemptions sells at whatever price clears. That pushes price away from where two-sided trading would have left it, and when the seller is done, it drifts back. You are being paid for supplying liquidity at an inconvenient moment.

Market makers have to reverse. The people quoting both sides of the market end up holding whatever the flow hands them and must unwind it. Their unwinding is mean-reverting pressure, mechanically.

It has a high win rate, which is psychologically the reverse of trend following. Most trades are small winners. That makes it easy to trade and dangerous to size — see below.

Say the cost

Mean reversion wins often and loses badly. It is the family most likely to show a beautiful equity curve for two years and then give all of it back in a week, because the losing case is not a small loss — it is the one time the move was information. The curve's smoothness is not evidence of safety; it is what this payoff shape looks like right up until it isn't.

What it looks like¶

A price stretched well below the day's volume-weighted average, then closing back toward it. vwap_reversion takes one number:

Parameter	Means
`entry_deviations`	how far below VWAP a close must be to count as stretched

flowchart TD
    B["5-minute bar closes"] --> Q{"Close more than<br/>entry_deviations<br/>below VWAP?"}
    Q -->|yes| G["Ask the risk gate"]
    Q -->|no| H{"Holding?"}
    G --> H
    H -->|yes| X{"Back at VWAP?"}
    X -->|yes| S["Exit"]
    X -->|no| W["Hold"]
    H -->|no| W

The exit is the mean itself, which is the defining feature of the family: the target is known when you enter. A trend follower has no idea where its exit will be; a mean reverter knows exactly, and that certainty is what caps the upside.

This is a 5-minute rule — VWAP resets each session, so it is meaningless on daily bars, and Arvo refuses to run it on them. It also needs at least five bars of a session before a volume-weighted deviation means anything.

When it fails¶

It sells the start of every trend. In a market that keeps making new lows, "far below VWAP" is true every bar, all the way down. Each trade is entered a little lower and each one loses. This is the mirror of a trend follower in a range, and it is the same mistake with the sign flipped.

The payoff is bounded up and unbounded down. Your profit is capped at the distance back to VWAP. Your loss is capped at nothing in particular. A win rate of 75% with that asymmetry can still have negative expectancy — go back to the expectancy table and put the numbers in.

A stop is not optional here, and it fights the strategy. The premise is that price has gone too far, so any stop is a bet against your own thesis at the point it is most stretched. Trading without one is worse. There is no comfortable answer, and anyone who tells you otherwise is selling something.

Intraday means costs dominate. Many trades, each with a small target, all paying the spread twice. This is the family where the cost tiers do the most work.

Test it¶

Intraday rules need 5-minute bars in your library, so fetch them before you run.

rulesets/stretch.json:

{
  "name": "stretch",
  "label": "VWAP reversion",
  "premise": "A close well below the session's VWAP returns to it often enough to pay for the days it does not.",
  "interval": { "step": 5, "unit": "minute" },
  "kind": {
    "kind": "grid",
    "rule": "vwap_reversion",
    "fixed": { "trade_size": 10.0 },
    "axes": { "entry_deviations": [1.0, 1.5, 2.0] }
  }
}

What to look at, in this order:

The trade count will not be your problem here — an intraday rule generates plenty. That makes this the best family to learn on, because you will get real verdicts rather than Inconclusive.
Compare stated against conservative costs. This is the test that matters for this family. A rule earning a few basis points per trade and paying the spread twice is usually a cost story.
Look at the worst trade, not the average. On the Trades tab, find the largest loss and ask whether your risk_per_trade would have survived several of them in a row. The average trade is not what ends accounts.
Check the regime split. Expect this rule to earn in ranging periods and bleed in trending-down ones. If it does, you have confirmed the theory and learned what you own.

Watch¶

(to be chosen) — VWAP, how it is computed and why institutions watch it.
(to be chosen) — a high-win-rate strategy failing, with the trade sequence.

Next¶

Relative strength — ranking names against each other instead of judging each on its own.