The 'Bear-Market' Strategy That Loses Principal in Bear Markets — Specialized for Survival, Not Profit

A trading strategy calls itself 'bear-market-specialized.' When the market breaks, it picks the names that have fallen too far and rides the brief bounce back up to recover principal (internal codename: deadcat-v1). The name is plausible. If a bear market is frightening because principal melts away, then a strategy built to aim at that bear market and win the principal back sounds like exactly the thing you need.

You have to hold the strategy's vocabulary in hand before the numbers below make sense.

• Oversold: fallen too far, too fast in a short span — a rubber band stretched to one side as far as it will go. The strategy reads the stored snapback force as an entry signal. But if the band snaps (the fall continues), there is no snapback.
• Dead-cat bounce: a brief upward jump partway down a falling trend. The phrase comes from the cynic's line that "even a dead cat bounces if it falls from high enough." The strategy eats this bounce to recover principal. It must be told apart from a genuine recovery in which the trend itself turns over (a V-shape) — and this distinction decides everything that follows.
• R-multiple: the realized profit or loss on a single trade, divided by the amount you can lose on that trade. Here the amount you can lose is not the full cost but the part exposed at the stop-loss line — 4% of cost. That is taken as 1R. Buy 1,000,000 won's worth and set the stop 40,000 won below, and that 40,000 won is 1R. Come out having made 80,000 won and it is +2R; get cut exactly at the stop and it is −1R. Because the ticker and the price level differ from regime to regime, won-denominated P&L can't be compared across them; you have to convert to "how many multiples of the 1R you staked did you win or lose" before regimes will line up for comparison.

The first time this strategy was run against the market, it fell well short of the bar. Over the measurement window, the KOSPI rose +184.32%. Just buying and holding over the same window compounds +142.54% on the sample basis.The sample's buy-and-hold (+142.54%) coming in below the index (+184.32%) is only because the index and the trading sample are composed differently; measure it either way, and the market more than doubled. The strategy, meanwhile, managed only +45.03%.

Two promotion criteria had been fixed in advance: beat the market, and keep maximum drawdown (the deepest fall from a peak) within 15%. The strategy missed both marks. Its return trailed the market by roughly three times, and its maximum drawdown, at 22.51%, breached the 15% cap. The verdict was failure.

This is usually where the story ends. A failed backtest (a simulation that runs a strategy over historical data) means a bad strategy, one to discard. But this failure had one strange thing about it. A strategy that earns in bear markets had been measured, of all places, in a surging market where the KOSPI jumped +184%.

Run the same strategy through four different regimes. There are three regime types (a rising market, a market that rebounded after a crash, and a market that slid down without rebounding), but the sliding-down type spans two years, so there are four report cards. The question that sorts the signs of those four cards narrows to just one. Did the rebound come?

The First Report Card Was a Misdiagnosis

Start by correcting what that verdict of "failure" actually is. Drop a bear-market strategy into a KOSPI +184% surge and losing is a foregone conclusion by design. A strategy that targets oversold bounces to take a small bite and get out has no way to beat a market that more than doubles just by being bought and held. This is not a signal that the strategy is wrong. It is a signal that this ruler can't measure it.

In an earlier piece I wrote that even the same semiconductor takes a different value depending on which ruler you measure it by (다음 주 코스피, 실물로 읽으려면 — 7/21 순통관과 7/22 PPI는 같은 반도체를 다른 자로 잰다). This time it is not the object but the strategy. Even the same strategy has the sign of its result decided by which regime you measure it in.

Instead of quietly burying the failure, split the original strategy quarter by quarter and a regularity surfaces. The P&L was aligned with the regime. In the quarter where a rebound followed a crash (Q1 2026, when the KOSPI took a −19.89% correction mid-quarter and then recovered), it earned heavily, +30.32% on capital. By contrast, in the rising market that pushed up without a correction (Q2, KOSPI +54.72%) and the falling market that slid without rebounding (Q3, KOSPI −17.86%), it lost in each.

But this is an observation that gave rise to the hypothesis, not evidence that proved it. The drawdown threshold that separated the quarters was just a conventional value, not a rule nailed down in advance. The observation only frames the question. Does this strategy's performance really split by regime type?

If so, the question has to be tested with a rule nailed down before the results are seen. And for that, the ruler had to change first — because the single surge window that produced the failure held only one falling quarter (Q3) with which to adjudicate the regime hypothesis.

Changing the ruler comes at a price. To secure several regimes you have to widen the sample out to 2018, 2020, 2022, and 2025 — but the original strategy runs on 60-minute bars (data at 60-minute intervals), and historical data that fine is hard to obtain going far enough back. So the multi-regime test is an analogous strategy run on daily bars (one-day intervals). Change the bar and the meaning of the indicators changes too, so strictly it is a different strategy.The lookback window of the representative indicator that measures the oversold condition in the original strategy is a few hours on 60-minute bars, but widens to 14 trading days on daily bars.

Here I have to be honest. This test looks not at the original strategy's performance but at a single principle: it works in regimes where the rebound comes, and it leaks in regimes where it doesn't. Because it is on daily bars, the absolute P&L can't be set side by side with the original strategy's 60-minute figures; what matters here is only the relative comparison across regimes.The decision and execution engine uses the same code as the original strategy verbatim, with only the bar interval changed. So the absolute magnitudes go unverified, and the promotion verdict's failure stands as well. I do not assert that this pattern observed on daily bars is a principle independent of bar interval; I run the relative comparison only on the assumption that it is a principle.

The Order of the Four Report Cards

Now spread out the four report cards obtained by changing the ruler. The character of each regime was not hand-picked but labeled from measured KOSPI index moves. Hit a bottom and fully recover within the same year, and it is a crash-then-V-bounce. Fail to recover within the year after the bottom, and it is a sustained grind (a slow, dragging decline). A year that rose without a correction was set as rising (the baseline).

Read the order. The regime this strategy earns in best is not a bear market but a rising market (+0.690R). Among bear markets, 2020 — which came back in a V — was thinly positive (+0.148R), while 2018 and 2022, which slid without rebounding, bled into the negative. One thing to flag: that 2020 crashed −35.7% in March, fully retraced within the year, and finished the year up +32.1%. The sole basis for the V-bounce is in fact a net-up year, so this thin positive is thin at its very foundation.

Here the signboard flips for the first time. "Bear-market-specialized" is not profit-specialized but survival-specialized. It doesn't make money in bear markets; it feeds well in rises and corrections and, in a grind, tries to die less (and even fails at that). And even that thin positive attached only in the regime where a rebound actually followed the crash.

Why lose in a grind? Break down the win rate by regime and the mechanism shows. Here the win rate is nothing other than the rebound-arrival rate — the share of trades that, after entry, closed for a profit because the rebound actually came. This value steps down from 54.5% in the rising market to 40.5% in the V-bounce, and again to 33.3% and 25.8% in the two grinding regimes. It overlaps exactly with the sign order of the average R. The less the rebound arrived in a regime, the worse the result.

The same pattern shows in the stop overshoot. A stop overshoot is when you fail to sell at the stop-loss line (cost −4%) and get cut lower, swept up in a gap or a plunge. It is like hitting the brakes on ice and sliding past the stop line. But the overshoot itself does not sort the regimes — it is a background constant that occurs evenly across all of them.The rate of overshoot per losing trade is an even 85–100% across every regime. Overshoot is a value set by a stock's volatility, not the regime, so it has no power to discriminate between regimes. What sorts the regimes is not whether an overshoot occurs but whether the rebound comes in the first place, so the trade never reaches the stop at all. It enters on an oversold signal, but the rebound doesn't come. It falls further, and it gets cut even past the stop-loss line. This is the measured version of the trading proverb about repeatedly catching a falling knife.

The signboard flips a second time. This strategy loses principal in exactly the regime it claims to defend.

Up to here, honestly, none of this is very surprising. "Don't catch a falling knife" is something every trader knows, and expecting a strategy that enters on oversold signals alone to get hurt in a sustained decline is close to common sense. The real thing to ask comes next: how much can this diagnosis be trusted? A backtest is an easy thing to knead in your own favor, and the negative numbers just seen could as easily be artifacts of bias.

Before that, one thing to nail down. This negative sign is no accident. The test carried a falsification condition set in advance: if the grinding regimes' average R came out at or above 0, or equal to or better than the V regime, the hypothesis that "the regime sorts the sign" breaks on the spot. The measurement put both grinds below 0 (−0.413R and −0.661R) and the V above them (+0.148R), so the falsification condition did not trigger. But this negative speaks to the sign and order across regimes, not to the magnitude of the absolute loss.−0.661R does not mean principal is wiped out; it means that for every 1R staked, you come out having lost roughly two-thirds of it. A serious bleed, but not an emptied account.

So far the two lenses have asked which ruler you measure the same strategy with. That a surging market is a ruler that can't measure it; that the regime is a ruler that sorts the sign — that is how the strategy's identity was established. Up to here is "what is true." From now the three remaining lenses are not about changing the ruler but about how much the measured numbers can be trusted. They grade these facts.

Grade One: A Guarantee It Wasn't Fitted After the Fact

Medicine has a practice called clinical-trial pre-registration. Before a drug is ever given to people, "what will count as success" is fixed in advance in a public registry (the US made it a legal requirement in 2007). The point is to seal off the self-deception of swapping in a favorable criterion after seeing the results.What statistics calls "p-hacking" — kneading the data this way and that and reporting a significance that surfaced by chance as if it were real.

This multi-regime test was designed exactly that way. The decision rule was nailed down in a document before the results were seen: the V-bounce average R must exceed the maximum of the two grinding regimes, and both grinding regimes must be at or below 0. How to select the regimes, what to label their character, how many trades to require at minimum, how to cut the periods — all of it was set before the results. Change the rule later and it stays on the version-control record for all to see.

Then comes this objection: "The 20-trade-per-regime threshold is low, and only four regimes were chosen. Pick them differently and you'd get a different result, wouldn't you?" The guards are layered. The character was not hand-picked but labeled from measured KOSPI index moves (V = full recovery 94 trading days after the bottom; grind = no recovery within the year). And the period was not carved down to the falling stretch alone but fixed to the whole year (January–December). The moment you pick the window's boundaries by hand is precisely where room opens for the researcher to knead the result, so that door was shut. The threshold was 20 trades, but the actually closed trades ran to 37, 62, 72, and 44 per regime, all clearing it. It is exactly how clinical pre-registration blocks p-hacking.Still, it is not a flawless pre-registration. The original strategy failed promotion on 60-minute bars, after which the ruler was changed to daily bars and this multi-regime test was run — so the arena of judgment was moved after a bad result had come in, which leaves room for goalposts set after the fact. The sample is also only dozens of trades per regime, so statistical significance was not separately tested (N=4).

That does not mean "survival-specialized" is proven wholesale. What the pre-registration supported is narrow: "V-bounce average R > grinding average R, and both grinds at or below 0." That means the sign and order are right, not that each number's magnitude is robust. In bar-chart terms, the ranking of the bars is correct and only the heights of the bars wobble. That the V edge is thin at +0.148R, and that this thin positive is vulnerable to bias and cost, both lie outside the pre-registration rule. The pre-registration nailed down only the skeleton (it loses in a grind); what remains is to grade the robustness of each sign separately.

That grading yields an unexpected result. The two grading items — survivorship bias and cost — do not shake this strategy's two conclusions equally. We take up which one each touches, and how, section by section.

Grade Two: The Bombers That Came Back

Why judging only by what survived leads you astray — statistics already solved this on the bombers of the Second World War. To flag it up front: the survivorship "bias" meant here is an entirely different phrase from calling this strategy "survival-specialized" earlier. This one is the statistical trap that arises when only the survivors remain in the sample.

Line up the bombers that came back from combat and plot the distribution of bullet holes, and intuition says reinforce the places with many holes (the wings, the fuselage). The statistician Abraham WaldThen with the Statistical Research Group (SRG) near Columbia University. answered the opposite: reinforce the places with no holes. A plane that came back means it was hit there and returned anyway. The truly fatal areas are the ones missing wholesale from the sample because the planes hit there could not return — that is, the places where no holes show. Wald's method was to back-calculate the vulnerability of what did not return from the damage distribution of what came back alive. His lesson reduces to one sentence: interrogate the absence.

The stocks this strategy's backtest looks at are exactly those "bombers that came back." The pool of tradable names (the universe) is the stocks that survived to today and rank in the top 350 by liquidity. Whichever past year you measure, you are in effect trading not that year's market but only "the names that grew into today's winners."

This is called look-ahead selection bias, because it is as if the sample were picked with knowledge of the future.

Here the skeptical reader shoots back: "If the universe is biased like that, then the average R by regime is all untrustworthy, and the grinding negatives are artifacts of the bias too, no?"

First a tempting but wrong defense has to be cleared away: "the bias runs optimistic, so the negative conclusion only gets firmer."If dead stocks were dropped, then the down years' performance would be recorded better than reality, and if the grind is still negative, the truth must be worse — a plausible inference. This strategy's data disproves that defense. The stocks missing from the universe did not vanish by going under; they are companies that in that year were not yet even listed. They grew later and made it into today's top 350 (a stock that survived to today cannot, by definition, be delisted). So "the worse the down year, the worse the bias" does not hold. Because the stocks missing in 2022, a bear-market year, are in fact fewer than in 2020, a net-up year.

Look-ahead selection bias is real. But whether it pushes the thin V edge up or down — that direction — has not yet been measured for this strategy. Whether trading only today's winners missed the rebound fuel of dying stocks, or dodged the overshoot losses of such stocks, cannot be known. Here Wald's lesson comes back on itself. The skepticism that looking only at the planes that came back leads you astray must be applied to this very piece, which is the one selling that skepticism. Apply Wald's answer — look to what's missing — honestly to this data, and the answer that comes out is "the direction has not yet been measured."

So survivorship bias is not a leg holding up any conclusion in this piece. Strip the bias away and the negative conclusion could get worse or less bad. The thin V edge could vanish or remain. Neither can be asserted from this data. Survivorship bias is not a leg holding up a conclusion but an honest limitation that bears equally on both sides.

Grade Three: The Sign Hangs on a Decimal Place

The third grading item is transaction cost. It is an entirely different axis from the pre-registration above, but it pushes the conclusion in the same direction.

A backtest's performance is a figure that counts trading as free. In reality, every buy and sell bleeds out the gap between the bid and the ask (the spread). In the original strategy's promotion-decision run, raising this round-trip cost from just 0.10% to 0.40% flips the sign of the realized net P&L. 0.40% is the same order of magnitude as the actual spread in the Korean market. A P&L that was about +1.38 million won at a 0.10% round trip becomes about −250,000 won at 0.40%.

A caution. This sign flip is a number from the 60-minute original strategy's surge-market run, so — because of the bar difference noted earlier — it can't be set side by side with the daily-bar V edge (+0.148R). So I do not say "cost flips the V edge." What can be said is this: the very claim that "this strategy makes money" hangs on the second decimal place of the round-trip cost. A performance said to be positive crosses into negative on one realistic transaction-cost assumption.

This is not a fluke unique to our backtest. Academic finance has independently confirmed the same direction. A 2010 paper showed that applying real quotes makes much of the positive return of short-term reversal trades disappear. The direction is the same as ours.That is, the return is an artifact not of a real liquidity recovery but of quote bounce (the illusion generated by prices ricocheting between the bid and the ask).

Here a last doubt remains: "If the sign crosses over on cost, then the positives are all a mirage, and if you earn nothing, isn't even 'survival-specialized' meaningless?" But what cost kills is only the positive figures. Lay cost on top of a negative (the grinding loss) and the sign stays put; only the number gets worse. So the cost vulnerability actually pushes this piece's thesis — toward the conclusion that "it is not profit-specialized."

The three grades are done. Pre-registration nailed down the sign and order; survivorship bias, its direction unknown, stayed an honest limitation; cost threatened only the positive figures. Now it is time to lay these three over the four report cards.

Rock and Mirage

Lay the four report cards and the three grades over one another, and this strategy's two claims reveal themselves to be of entirely different material.

• The negative conclusion — that is, "it loses principal in a sustained grinding decline." Pre-registration nailed the sign and order before the results were seen, holding it up substantively. Lay round-trip cost on top and the negative's sign does not change. Here cost is not actively holding this conclusion up; add cost to a negative and it stays negative, so it simply was never a threat in the first place. So the negative conclusion clears both hurdles at once: the substantive support of pre-registration, and the fact that cost can't shake it. The positive conclusion passes neither of the two. These two legs stand even without knowing the direction of survivorship bias. The direction of the bias itself is still unmeasured and an open question, and this piece honestly leaves that direction at the end as a falsifiable stake. So however you set the sample threshold, this side — having passed both skepticisms — is the relatively surviving rock.
• The positive conclusion — that is, "it earns in a bear market (the V-bounce)." The reasons it is thin are layered. The magnitude itself is thin at +0.148R; its sole basis, 2020, is in fact a net-up year, so the basis is thin too; and a survivorship bias of unknown direction is laid on as one more. On top of that, this strategy's positive realized P&L generally hangs on the decimals of the round-trip cost. It is a mirage that vanishes when you change the ruler.

The fact that a backtest carries survivorship bias commonly leads to the symmetric skepticism of "so none of it can be trusted." On the other side is the easy inversion of "the bias runs optimistic, so the bad conclusion is if anything trustworthy." This piece passes both. The direction of the bias was not measured, so neither side can lean on the bias. And yet the reason the two conclusions are not of the same material is that skepticism does not gnaw at them equally. The negative conclusion stands on two legs that hold even without knowing the bias's direction — pre-registration and the cost arithmetic — while the positive conclusion, unprotected by those two legs, rests on thinness and a fee rate. So skepticism sorts the two conclusions — which is rock and which is mirage. Just as Wald read "where to reinforce" from the distribution of bullet holes, what matters here is not the direction of the bias but which conclusion stands even without knowing the bias.

Two reservations are carried along. First, this verdict that "it loses principal in a grind" is a principle-level conclusion observed in the daily-bar analogous strategy — how the 60-minute original strategy actually fares in a sustained grind has not yet been measured directly.The direct evidence for that amounts to a single quarter, Q3 2026, and even that is a post-hoc observation, not a pre-registered verdict. So this conclusion should be read not as a settled verdict on the original strategy but as a principle that the sign splits by regime. Second, there are only four regimes, and of them only two are falling. So this axis — that "did the rebound come" sorts the sign — does not cleanly separate, on this data alone, from the market drift that "in a rising year you earn whatever you buy" (the V-bounce and the rising were up years; the two grinds were down years). But one thing suggests the axis is not annual direction alone. Even within the same rising year (2026), the quarter where a rebound followed a crash earned +30.32%, while the quarter that slid without rebounding lost −3.65%.But this contrast, too, does not fully eliminate the confounding. That earning rebound quarter (Q1 2026) is a rising leg that recovered after a correction, so it is also a stretch that locally carried positive (+) beta — the confounding was merely moved from annual to quarterly. Being an observation within a single year rather than a pre-registered verdict, it is not decisive. So the statement that "the presence of a rebound is the axis" is carried as a reservation, not an assertion. Even if the negative conclusion is a rock, the thinness of the sample remains — the grinding regimes propping up that rock are themselves only two (2018 and 2022).

Then the question remains of what practical use this is. "Whether the present is a V-shaped decline or a grinding one is knowable only after the fact, and if you can't tell them apart in advance, isn't this conclusion useless?" An accurate point — and that is precisely this piece's implication. Telling in advance whether the present decline is V-shaped or grinding is not this strategy's job. From its very specification, the strategy nails it down that reading the market is left to the user. That is what makes the "bear-market-specialized" signboard dangerous. Flipping the bear-market switch does not make every bear market the same one, yet the signboard speaks as if it defends every bear market. What this strategy actually defends is only the bear market where a rebound follows the crash; in the bear market that slides without rebounding, it instead eats away at principal.

So "bear-market-specialized" is not an honest name. To be precise, it is crash-then-V-bounce-specialized, and outside that it is counterproductive. And the strategy will not make the very judgment the would-be user needs — which bear market is this — for them. The promotion verdict's failure is still a failure. This piece did not reverse that failure; it only clarified what kind of failure that was.

Finally, I place my stake. I hold that this strategy's only robust conclusion is the negative one (it loses principal in a sustained grinding decline). Suppose the two remaining tasks were done — reconstructing a point-in-time universe to strip out look-ahead selection entirely, and putting in transaction cost at a real account's quotes. If the grinding regime's average R then came out at or above 0, I am wrong, and this strategy is more robust than I have said.