Intermediate5-8 min readTopic 8 of 20

    Trend Following vs Mean Reversion

    Rohit Singh

    Mr. Chartist · SEBI RA

    Module Progress
    0/20
    Module

    Almost every systematic strategy is one of two bets wearing different clothes. Trend following buys strength and sells weakness, betting that what has moved will keep moving. Mean reversion sells strength and buys weakness, betting that what has stretched will come back. They are not two answers to the same question — they are opposite answers, and each one is at its most confident exactly where the other is bleeding. Choosing between them is not a question about which is better. It is a question about which way of being wrong you can live with.

    One price path through contraction, expansion and contraction again. The circles are where a mean-reversion rule acts; the squares are where a trend rule acts; the crosses are where each one runs into the other's home condition.A single price line moves through three conditions: a sideways range, a directional expansion out of it, and a second range at a higher level. A mean-reversion rule acts at the edges of the two ranges, buying the lower edge and selling the upper edge. A trend rule acts only when price leaves a range and keeps going. Two clash points are marked: the fade rule selling into the expansion, and the breakout rule buying an edge inside the upper range that price immediately returns from.ONE PRICE PATH — THREE CONDITIONS, TWO OPPOSITE INSTRUCTIONSCONTRACTIONEXPANSIONCONTRACTION AGAINMean reversion acts at the edges — it needs price to come backTrend following acts on the leave — it needs price not to come backEach family's clash point is the other family's home condition
    One price path through contraction, expansion and contraction again. The circles are where a mean-reversion rule acts; the squares are where a trend rule acts; the crosses are where each one runs into the other's home condition.

    What is trend following, in one paragraph?

    A trend-following rule waits for price to leave a region it has been trading in, and then acts in the direction it left. The signal is the departure itself. A breakout above a base that has held for thirty-odd candles, a close beyond the highest high of the last N candles, a series of higher highs and higher lows resuming after a pullback — these are all the same instruction stated differently: something has already happened, join it.

    The defining feature is that the rule has no opinion about value. It does not think the price is cheap. It thinks the price is moving, and it is willing to buy something that has already risen precisely because it has already risen.

    What is mean reversion, in one paragraph?

    A mean-reversion rule waits for price to stretch away from where it has been spending its time, and then acts against the stretch. The signal is the distance. A long lower wick at the bottom edge of a range that has contained price for fifty candles, a fast push into an area price has been rejected from repeatedly, a gap that opens far from the prior session's body and starts filling — again, one instruction: something has gone too far, take the other side.

    Its defining feature is the mirror of the other. It has no opinion about direction. It thinks the current price is temporary.

    Why is this the same decision, twice?

    Notice what happens when you put both rules on the same chart. The exact bar that triggers a trend entry — a decisive close beyond the edge of a range — is the bar a mean-reversion rule reads as an extreme worth fading. They are looking at the identical price event and reaching opposite conclusions.

    That is not a flaw in either. It is the reason both can exist. A market needs participants on both sides of the question, and the answer changes depending on what the market is doing at that moment — which is the whole difficulty, because you find out what it was doing several candles after it stopped doing it.

    Note — This is why 'which strategy is better' has no answer, and why anyone offering one is selling something. The question is incomplete without an instrument, a timeframe, a market condition, a cost assumption and a position size.

    How do the two families compare structurally?

    Read that table as two internally consistent packages, not as a scorecard. Nothing in either column is a claim about which one earns more. Both columns describe how each family behaves; neither describes how it performs.

    Trend followingMean reversion
    The betWhat is moving will keep movingWhat has stretched will come back
    Acts onThe departure from a rangeThe distance from a middle
    Entry feelsUncomfortable — you buy after a riseComfortable — you buy after a fall
    Typical outcome patternMany small losses, few large gainsMany small gains, few large losses
    Where the result comes fromA small number of decisions that ran a long wayThe accumulation of many ordinary decisions
    Exit logicOpen-ended — let it run until structure breaksBounded — a target near the middle it was stretched from
    Home conditionExpansion — range left behind, held for many candlesContraction — price keeps returning to a middle
    Killed byRepeated breaks that reverse within a few candlesOne break that does not reverse at all
    Hardest psychological demandSitting through a long run of small losses without changing the rulesCutting a losing position that has always come back before
    Sensitivity to costsLower per decision — fewer, larger moves carry itHigher — small edges are easily consumed by charges and slippage

    Why do the return distributions matter more than the averages?

    This is the part most comparisons skip, and it is the part that decides who survives.

    If you lined up every decision a trend-following rule took and grouped them by outcome, the shape would be lopsided in a specific way: a dense cluster of small negative results, because most breakouts stop and the rule cuts them, and a thin tail stretching a long way into the positive, because a handful ran. Remove that tail and there is nothing left.

    A mean-reversion rule produces the mirror image. A dense cluster of small positive results, because price usually does come back, and a thin tail stretching a long way into the negative, because occasionally it does not and the position was against a move that kept going.

    The two shapes, drawn deliberately without a scale. Neither axis carries a number, because these are qualitative shapes and not a claim about any strategy's results.Two unnumbered histograms. The trend-following shape has most of its outcomes bunched just on the losing side of break-even and a long thin tail extending into gains. The mean-reversion shape is the mirror image: most outcomes bunched just on the winning side of break-even, with a long thin tail extending into losses. Neither axis carries a number, because the shapes are qualitative and no performance figure is claimed.TREND FOLLOWING — SHAPE ONLYMEAN REVERSION — SHAPE ONLYBREAK-EVENBREAK-EVENMany small losses, a thin tail of large gainsMost decisions lose a little. A few carry the whole result.Many small gains, a thin tail of large lossesMost decisions win a little. A few undo a long run of them.Deliberately unnumbered. These are shapes, not results, and neither is a claim about outcomes.
    The two shapes, drawn deliberately without a scale. Neither axis carries a number, because these are qualitative shapes and not a claim about any strategy's results.

    What does the shape imply about how you experience each one?

    • Trend following spends most of its calendar losing quietly and a small part of it gaining loudly. The waiting is the strategy, not an interruption to it.
    • Mean reversion spends most of its calendar gaining quietly, which feels like competence, until the one occasion that does not revert arrives on top of a position built while it felt safe.
    • A trend follower who stops after a long losing run has thrown away the only decisions that were going to pay for it — and there is no way to know in advance which ones those were.
    • A mean-reversion trader who adds to a losing position because it has always come back is doing the one thing that converts the thin left tail into an account-ending event.
    • Neither shape is safer. One front-loads the pain and the other back-loads it, and back-loaded pain is far easier to underestimate while you are being paid.

    Watch out — Do not translate 'many small gains' into 'high win rate' and then into 'better'. A rule can be right on most of its decisions and still lose money overall, and a rule can be wrong on most of them and not. Any comparison that leads with a percentage of winners has told you the least useful number available.

    How does the arithmetic of the two shapes work?

    Strip it to a toy. Take a rule where each loss costs ₹1,000 and each win makes ₹5,000, and run twenty decisions of which four win. Sixteen losses is ₹16,000 out; four wins is ₹20,000 in. Now take the mirror: each win makes ₹1,000, each loss costs ₹5,000, and sixteen of twenty win. ₹16,000 in, ₹10,000 out.

    Both are positive here only because the numbers were chosen to be. Change the mirror case to three losses instead of four and it turns negative, on a rule that was still right 85 per cent of the time. That fragility is the point: in the mean-reversion shape, the whole result sits in the size and frequency of the rare loss, and that is exactly the number nobody can observe enough of.

    Result = (wins × average win) − (losses × average loss) − total costs
    • Neither term is meaningful alone. A high count of wins says nothing until the average sizes are attached.
    • Total costs — brokerage, exchange charges, STT, stamp duty, GST and slippage, applied per decision. More decisions means this term grows faster than the others.
    • Average loss — the term that is least reliably estimated, because the losses that matter most are the rarest. This is where a mean-reversion rule is most vulnerable to a small sample.

    Example — Illustrative arithmetic with invented round numbers. It is not a backtest, not a result, and not attached to any instrument or period. It exists only to show why the count of winners cannot be read on its own.

    What is regime dependence, and why can you not simply switch?

    A market condition — the regime — is the structural behaviour price is showing over a stretch of candles. Described without indicators, there are two that matter here. In expansion, price leaves an area and holds the new one for many candles; ranges break and do not come back. In contraction, price keeps returning to a middle; edges hold, and moves beyond them reverse within a few candles.

    Each family is built for one of them. That is not a weakness you can engineer away — it is what the family is.

    Each family aligned with one condition and fighting the other. Both keep trading in the condition that does not suit them, because neither can tell which one it is in until several candles after it changed.A two-by-two grid. Rows are the market condition: expansion, where price leaves a range and holds the move for many candles, and contraction, where price keeps returning to a middle. Columns are the strategy family. Trend following is aligned with expansion and fighting contraction. Mean reversion is aligned with contraction and fighting expansion. Neither cell is empty, because both families keep trading in the condition that does not suit them.NEITHER FAMILY STOPS TRADING WHEN THE CONDITION TURNS AGAINST ITTrend followingMean reversionEXPANSIONrange left behind, held for many candlesAlignedprice leaving and not returning is the whole thesisFighting itevery 'too far' level is passed and then passed againCONTRACTIONprice keeps coming back to a middleFighting iteach break reverses inside a few candlesAlignedprice returning to a middle is the whole thesisYou cannot know which row you are in until several candles after it changed. That lag is the cost of both.
    Each family aligned with one condition and fighting the other. Both keep trading in the condition that does not suit them, because neither can tell which one it is in until several candles after it changed.

    Why is a regime filter harder than it sounds?

    1. 1

      Regimes are only labelled in hindsight

      You can point at a stretch of NSE chart and say it was trending. You are reading the whole stretch at once. Live, you have only the left-hand side, and the candles that would confirm the label have not printed.

    2. 2

      Any filter has a lag, and the lag has a cost

      A filter that needs twenty candles to declare a trend has, by construction, missed twenty candles of it. Shorten it and it flips regime on noise; lengthen it and it arrives after the move. There is no setting that removes this — you are only choosing which error to make.

    3. 3

      The transitions are where both families lose

      The candles where a range is becoming a trend, or a trend is decaying into a range, belong to neither family cleanly. A filter does not eliminate that zone. It just decides which side of it you are wrong on.

    4. 4

      Running both is a real answer, and not a free one

      Two rules with opposite exposures can hold a portfolio steadier through a regime change than either alone. The cost is that one of them is visibly losing at all times, and you have to keep funding and running the one that currently looks foolish. Most people cannot.

    5. 5

      A filter is another thing that can be overfitted

      The moment you add a regime switch, you have added parameters, and parameters can be tuned until the switch happens to have fired correctly in your sample. Walk-forward validation is covered separately in this module, and it applies to the filter as hard as to the strategy.

    Why are the exits as opposite as the entries?

    The entries get all the attention, but the exit is where the two families genuinely diverge, because each one's exit has to protect the shape it depends on.

    A trend-following exit must be open-ended. The rule cannot take a fixed target, because a fixed target amputates the thin tail that the whole family lives on — you would be cutting short exactly the few decisions that were supposed to pay for all the small losses. So the exit is structural: hold until the sequence that defined the trend breaks, and accept giving back part of the move every single time.

    A mean-reversion exit must be bounded, for the opposite reason. The thesis was that price had stretched too far from a middle; once it is back near that middle, the reason for the position no longer exists. Holding on to see if it keeps going converts a mean-reversion position into an accidental trend position with none of the sizing or the exit logic a trend rule would have used.

    • A trend rule with a fixed target is not a trend rule. It has kept the uncomfortable entry and thrown away the payoff shape.
    • A mean-reversion rule with an open-ended exit has kept the comfortable entry and adopted the other family's need for follow-through.
    • Both families need the invalidation written before entry — the price at which the structural reason has gone, not an amount of money you feel able to lose.
    • A time-based exit is legitimate for mean reversion: if price has not come back within the number of candles the thesis assumed, the thesis is stale whether or not the level was hit.
    • The same time-based exit is usually wrong for trend following, where the whole edge is in the decisions that take an unusually long time to develop.

    Does the choice depend on the instrument and the timeframe?

    Heavily, and this is where a comparison in the abstract stops being useful.

    An index behaves differently from a single stock, because it averages away company-specific surprises and cannot gap on one set of quarterly numbers. A liquid large-cap behaves differently from a thin small-cap, where the spread you pay and the size you can move are themselves a strategy constraint. And the same instrument shows different structure at different scales — a stretch that is a clean trend on a weekly chart is a series of ranges and reversals on a 15-candle chart of the same period.

    So the honest statement is that a family is not chosen for a market. It is chosen for a specific instrument, on a specific timeframe, under a specific cost assumption — and changing any of those three can invalidate the choice entirely. That is also why a rule that looked robust in research fails when it is moved to another instrument that 'looked similar'.

    Pro tip — Whenever someone tells you a strategy family works or does not work, ask which instrument, which timeframe and what costs were assumed. If those three are missing, the statement is not testable and cannot be acted on.

    What breaks each family, specifically?

    Trend following — the chop
    A sequence of breaks that reverse within a handful of candles. Each one is a small planned loss, and the rule is behaving exactly as designed; the damage is cumulative and psychological. The invalidation is not a single price — it is a stretch of market where every departure from a range failed to hold.
    Trend following — the gap through the exit
    The exit level exists on the chart, not in the market. On an Indian stock, results, a board decision, a block deal or a rating change can open price past it. Position size, not the stop level, is the only real control over this.
    Mean reversion — the break that does not come back
    The single failure that matters. Price leaves the range and keeps going, against a position that was sized as if it would return. Everything the rule had accumulated is available to be handed back here.
    Mean reversion — averaging down
    Not a market condition; a decision. Adding as the position moves against you converts a bounded loss into an unbounded one, and it feels most reasonable precisely on the occasion it is fatal, because it has worked every previous time.
    Both — cost drag
    Charges and slippage are subtracted per decision, not per year. A high-frequency mean-reversion rule with a small per-decision edge can be arithmetically sound and still lose after brokerage, exchange charges, STT, stamp duty, GST and the spread it actually pays. Model costs before conviction.
    Both — the structural reason stopped applying
    The clean invalidation for either family is not a bad run. It is the observation that the market behaviour the rule depended on is no longer present — breaks holding, or edges holding. Strategy Decay and When to Switch Off deals with telling that apart from ordinary variance, which is genuinely hard.

    Which beliefs about these two families are simply wrong?

    "Mean reversion has a higher success rate, so it is better"
    The proportion of winning decisions is a property of the shape, not of the quality. It is high by construction, because the family takes small gains and rare large losses. Read on its own it tells you almost nothing about whether the rule is worth running.
    "Trend following only works in bull markets"
    The family needs expansion, not upward direction. A sustained decline is an expansion too, and a rule that can act in both directions is agnostic about which way price left the range. What it cannot survive is a market that keeps returning to where it started.
    "Mean reversion is just buying dips"
    Buying a dip inside an ongoing trend is a trend entry — you are joining a move on a pullback and you need the move to continue. Mean reversion needs price to reverse and return to a middle. The two look identical at the moment of entry and demand opposite things afterwards, which is exactly how positions get held for the wrong reason.
    "A regime filter solves the problem"
    It relocates it. Every filter lags, every filter can be tuned to fit the sample it was built on, and the transition candles between regimes belong to neither family. A filter changes which error you make; it does not remove the error.
    "You can combine both into one rule that always works"
    A rule that both buys breaks and fades extremes has to decide which it is doing on every bar, and that decision is the regime problem again with more parameters attached. Running two separate rules with clear, separate risk budgets is honest. Blending them into one is usually a way of hiding the choice rather than making it.

    How do you choose between them?

    Not by comparing performance, because you cannot verify anyone else's and your own sample is too small to settle it. Choose by asking which failure mode you can actually sit inside, because you will spend most of your time there.

    Answer these honestly before you write a line of code

    • Can you take a long, uninterrupted run of small losses without editing the rules? If not, the trend-following shape will remove you before its tail arrives.
    • Can you close a losing position that has come back every previous time? If not, the mean-reversion shape has a specific way of ending you.
    • Does your chosen family match the instrument? An index and a thin small-cap do not behave the same way, and neither does the same instrument on a 5-candle and a weekly chart.
    • Have you costed it per decision, with realistic slippage, before deciding it works?
    • Is your position size set from the invalidation level rather than from conviction? This matters more than the family you picked.
    • Do you have a written rule for switching the system off that you decided before you needed it?
    • Are you choosing based on how the equity path will feel, or on a number someone showed you? Only the first is verifiable by you.

    Note — This page is education, not advice. It does not recommend either strategy family, and nothing here is a suggestion to trade any instrument or run any system. Systematic trading carries market risk and engineering risk, and both can lose money.

    Key points

    Trend following bets that a move continues; mean reversion bets that it reverses. They are opposite answers to the same price event.
    Trend following takes many small losses for a thin tail of large gains. Mean reversion takes many small gains against a thin tail of large losses.
    The distributions are mirror images. Comparing average outcomes hides the thing that decides survival.
    A high proportion of winning decisions is not evidence of a better rule — it is a feature of the mean-reversion shape.
    Each family is aligned with one market condition and fighting the other, and cannot tell which one it is in until several candles later.
    Regime filters do not remove the problem; they choose which error you make, and they add parameters that can be overfitted.
    What breaks trend following is repeated breaks that reverse. What breaks mean reversion is one break that does not.
    Averaging down converts mean reversion's bounded loss into an unbounded one, and feels most reasonable when it is fatal.
    Costs are subtracted per decision. The higher-frequency family is far more exposed to charges and slippage.
    Choose by which failure mode you can sit inside, not by a performance number you cannot verify.

    Pro tip — Before you build either one, write down the worst stretch you are willing to sit through — as a number of consecutive losing decisions for a trend rule, or as a single loss size for a mean-reversion rule — and then size the position so that stretch is survivable rather than merely unpleasant. Do it before you see any results, because after you see them the number you write down will be the one your results allow rather than the one you can actually live with.

    Frequently asked questions

    Which is better, trend following or mean reversion?

    Neither, and the question is incomplete without an instrument, a timeframe, a market condition, a cost assumption and a position size. They are opposite bets that fail in each other's conditions. The useful question is which failure mode you can psychologically sit through — a long run of small losses, or a rare large one after a long run of small gains.

    What is the difference between momentum and mean reversion?

    Momentum, or trend following, acts in the direction price has already moved and needs the move to continue. Mean reversion acts against the move and needs price to return to where it had been spending its time. They read the same price event — a decisive break beyond a range — and reach opposite conclusions about it.

    Why does my trend-following system have so many losing trades?

    Because that is the shape of the family, not a defect. Most departures from a range stop, and the rule cuts them small by design. The result, if there is one, comes from a thin tail of decisions that ran a long way — which means cutting the system off during a losing run removes exactly the decisions that were going to pay for the rest.

    Is mean reversion safer than trend following?

    No — it is differently unsafe. Mean reversion produces many small gains and feels safe, and the risk is concentrated in the rare occasion when price leaves and does not come back. Trend following front-loads its discomfort as frequent small losses. One shape puts the pain first, the other puts it last, and back-loaded pain is far easier to underestimate.

    What is a market regime in trading?

    The structural behaviour price is showing over a stretch of candles. Described without indicators there are two that matter for this comparison: expansion, where price leaves an area and holds the new one for many candles, and contraction, where price keeps returning to a middle and moves beyond the edges reverse within a few candles. Trend following is built for the first, mean reversion for the second.

    Can I run both strategies at the same time?

    Yes, and holding two rules with opposite exposures can steady a portfolio through a regime change more than either alone. It is not free: one of the two will be visibly losing at any given time, and you have to keep running and funding the one that currently looks foolish. That is the part most people cannot maintain.

    How do I know which regime the market is in right now?

    You do not, with certainty, and any method that claims otherwise is describing hindsight. Every regime filter carries a lag: short settings flip on noise, long settings arrive after the move. You are choosing which error to make, not removing the problem, and the transition candles belong cleanly to neither family.