Hold the Dip: Why Buy-the-Dip Loses to Trend-Following EAs

The logic feels unbreakable: a 4% dip off an all-time high is a lower price than yesterday, so buying it must be the disciplined move. AQR put that instinct through 60 years of data — 196 buy-the-dip strategy variants on the S&P 500 from January 1965 to September 2025 — and the 2025 "Hold the Dip" paper returned a verdict most retail algo traders will not want to read: buy-the-dip underperforms simple passive accumulation more than 60% of the time, and it fails hardest in exactly the drawdowns it was supposed to exploit. For the MT4/MT5 developer choosing between a mean-reversion dip-buyer and a trend-following system as the core of an EA, this is not market commentary. It is an architecture decision with six decades of evidence attached.

The Seductive Math of Buying Cheaper

Buy-the-dip (BTD) is psychologically irresistible because it dresses a momentum bet up as value investing. You wait for a pullback, you buy "on sale," and the entry feels prudent rather than reckless. AQR's framing cuts straight through the comfort:

BTD is essentially value investing at a momentum horizon.

That sentence is the whole problem in eight words. The strategy tries to harvest a value premium — buy low — on a timescale where momentum, not value, dominates price behavior. AQR's second observation closes the trap:

The intuition that 'buying cheaper must be better' runs headlong into mathematical reality: you can't reliably identify dips in advance, and time out of the market has a price.

That price is the crux. To buy a dip you must hold cash waiting for one, and equities spend far more time rising than dipping. The S&P 500 set 1,325 all-time highs between 1950 and August 2025 — more than 17 per year on average (BNY Investments). A strategy whose precondition is "wait for a meaningful decline" is, by construction, sidelined during a large share of the market's upward drift.

What 196 Strategy Variants Actually Showed

The strength of the AQR study is its brute-force coverage. The 196 configurations span four dip depths, seven dip lengths, and seven holding periods — a grid wide enough that cherry-picking a flattering result becomes statistically honest rather than anecdotal. The aggregate findings:

• Underperformance frequency: BTD lagged a passive dollar-cost-averaging (DCA) benchmark more than 60% of the time across the 60-year window (AQR / Evidence Investor).
• Ending wealth: on average, BTD produced 18.7% lower ending wealth than simply DCA-ing into equities (AQR "Hold the Dip").
• Alpha: the average BTD variant generated just +0.5% per year of alpha versus passive equities, and only 8% of the 196 strategies showed statistically significant outperformance (Alpha Architect's analysis of the paper).
• Risk-adjusted return: BTD's Sharpe ratio came in 0.04 worse than passive overall; restricting to data from 1989 onward, the degradation widened to 0.27 — a 47% risk-adjusted penalty (Alpha Architect).

An 8% hit rate on statistical significance is the number EA developers should sit with. If you backtested 196 parameter sets of any system and only 8% cleared a significance threshold, you would correctly suspect that the survivors were noise, not edge. That is precisely the warning AQR is delivering about dip-buying as a category.

Where Buy-the-Dip Fails Worst: The Real Bear Market

The average underperformance is bad enough; the conditional failure is what destroys accounts. BTD's appeal rests entirely on fast V-shaped recoveries — and those are the exception, not the rule. The COVID crash of March 2020 (S&P roughly −34%, fully recovered in about five months) and the April 2025 "Liberation Day" snapback trained a generation of retail traders to buy every dip. But the representative cases are the 2000–2002 dot-com bust (−49%, roughly seven years to fully recover) and the 2007–2009 financial crisis (−57%, about 5.5 years to recover). In those, dip-buying means catching falling knives for years.

The numbers make the asymmetry stark. Across the four worst S&P 500 drawdowns since 2000 — each exceeding 20% — trend-following delivered an average +28.6% while the average BTD strategy lost 18.4% (AQR / Alpha Architect). That is a roughly 47-percentage-point swing in the precise environment where capital preservation matters most. Trend-following wins there because its core rule — exit and even reverse when price breaks down — is the structural opposite of adding to a losing long.

Why Dips Near All-Time Highs Are Too Shallow to Trade

There is a second, quieter problem with dip-buying near highs: the dips barely exist. Since April 2024, the S&P 500's average drawdown has been just 1.1%, and the largest dip available was only 4.2% off the all-time high (Investing.com analysis). A system that waits for a 5% or 10% pullback to deploy capital would have sat in cash for most of that span, watching the index grind higher without it.

And entering at the highs themselves has not been the trap intuition suggests. BNY Investments found that investing at all-time highs produced 5-year average returns of 10.5% versus 11.4% for all other days — a marginal gap, not a penalty severe enough to justify sitting out. The data quietly dismantles the premise that ATH is a dangerous entry that must be improved on by waiting for a discount.

The practical consequence for an EA is that near-ATH regimes starve a dip-buyer of setups while feeding a trend-follower a clean, persistent signal. Traders can size this scarcity for themselves by overlaying a drawdown-from-peak study on a live SPX chart — counting how often a tradeable dip of a given depth has actually appeared. Traders can monitor these levels in real time using TradingView, which provides drawdown and multi-timeframe overlays useful for measuring how shallow recent pullbacks have been before committing dip-depth thresholds to an EA.

The Trend-Following Counter-Record

The case is not merely that BTD is weak — it is that the alternative is demonstrably stronger across long samples. The SG Trend Index delivered a 4.7% annual alpha to equities with near-statistical significance, against BTD's 0.5% (Alpha Architect, referencing the SG Trend Index). Wesley Gray's summary was blunt:

Trend-following significantly outperformed, with the SG Trend Index delivering a 4.7% alpha to equities — far better than BTD's 0.5% average alpha.

The longer-horizon evidence is even more striking. A multi-century study of trend-following across 67 markets and four asset classes from 1880 to 2016 reported annualized net returns of 11.2% with a standard deviation of only 9.7% — roughly half the volatility of equities (cited by Alpha Architect). And Man AHL's trend program returned +82.5% from its January 2019 low through the start of the May 2024 drawdown phase, with an average recovery of +9.8% in the 12 months following a greater-than-10% rolling drawdown (Man Group, 2025).

None of this makes trend-following a free lunch. Man Group is candid that 2025's on-again, off-again US policy dynamics created a whipsaw that hurt trend programs — but frames it as transient:

Challenging phases for trend-following performance have frequently been followed by more favorable environments.

For a forex EA developer this whipsaw is not abstract: the DXY oscillated between roughly 98 and 107 across 2025 as tariff policy reversed, and that is exactly the chop in which a trend filter bleeds. The lesson is regime-conditional, not unconditional — which is the bridge to EA design.

The EA Translation: Win Rate Versus Expectancy

Reduced to the metrics EA developers already live by, the two architectures have opposite signatures:

• Mean-reversion / dip-buying EAs: typical win rate 70–85%, targets of 5–15 pips per trade — comfortable and frequently green in range conditions, but structurally exposed to trend extensions (MQL5 Articles / QuantifiedStrategies).
• Trend-following EAs: typical win rate 40–55%, targets of 50–100+ pips — lower frequency, more losing trades, but higher expectancy per trade when a trend actually develops (MQL5 Articles / ForexRobotLab).

The high win rate of the dip-buyer is the same seduction that shows up everywhere in this data. A 75%-win mean-reversion system that bleeds its entire equity in one un-reverted trend is the EA-level expression of BTD losing 18.4% in a bear market. The trend-follower's sub-coin-flip win rate is, conversely, the price of admission for the +28.6% it captured when it mattered. Expectancy times opportunity — not win rate — is the figure that survived AQR's 60-year grid.

// The asymmetry in one line of pseudo-logic
// Dip-buyer:  many small wins, rare unbounded loss
//   E[fail] = -1 * (full trend extension)
// Trend EA:   many small losses, rare large win
//   E[win]  = +N * (captured trend leg)
//
// AQR grid verdict: the second distribution
// compounded better across 1965-2025.

Building the Verdict Into Your EA Architecture

The honest reading of the evidence is not "never code mean reversion." Mean-reversion EAs are genuinely profitable in range-bound regimes — the S&P 500 itself spends more than 60% of its time in a 3%+ drawdown state and more than 50% in a 5%+ drawdown state (VITA Wealth Management / SlickCharts), and ranges are where fades earn their keep. The failure is regime-blind deployment. As the MQL5 community puts it:

The most common mistake is running a trend-following EA in a range-bound market or a mean-reversion EA during a strong trend — regime awareness is as important as the EA itself.

That reframes the build decision as a sequence rather than a binary:

1. Default to the trend architecture for the core. Across 60 years and four asset classes, the trend distribution compounded better and — critically — survived the deep drawdowns that destroy dip-buyers. If you build only one engine, the data favors this one.
2. Gate mean reversion behind a regime filter. Treat dip-buying as a conditional module that only arms when a regime detector confirms range conditions, not as the always-on core. An ADX or volatility-state filter is the difference between a profitable fade and an account-ending average-down.
3. Stress-test on a real bear sample. Validate any dip-buying logic on 2000–2002 and 2007–2009 data, not just the 2020–2025 V-shapes. If it does not survive a multi-year decline in backtest, it will not survive one live.
4. Size dip thresholds to the regime, not the wish. Near all-time highs, a 5%-dip trigger may fire once a year. Calibrate to the 1.1% average / 4.2% maximum reality rather than to the textbook 10% correction.

AQR's 196 variants did not prove that buying low is irrational — they proved that you cannot reliably identify the lows in advance, and that the time and tail risk spent trying cost more than they returned. For the EA developer, the cleanest expression of that finding is structural: make trend-following the spine of the system and let mean reversion earn its place as a regime-gated module. The market does not reward the comfort of a high win rate. It rewards the architecture that is still standing after the drawdown the backtest never showed you.