We Tested the Darvas Box Breakout on 17 Years of US and Thai Data

Personal research and trading notes — not investment advice. The author does not provide licensed advisory services.

A popular YouTube backtest of a weekly "Darvas-style" box breakout reported 195 trades, an average of +0.74R per trade, and a total return of +413% over 2022 to 2026, against +75% for the S&P 500 buy-and-hold. We took the published settings, froze them before looking at any outcome, and ran them across 17 years of US and Thai price history. This piece sets out what the data shows, and where it stops showing anything.

What the published claim was

The source is a video backtest on a weekly chart. Its on-screen settings: a consolidation box of at least 6 weeks with a range of no more than 20%; a close at least 4% above the box top; a close in the top quarter of the week's range; volume at least 1% above the prior week; a close above the 20-week EMA with the MACD line above its signal line. The stop sat 33% of the way up the box, capped at 16% risk. The reported result for 2022-01-01 to 2026-10-14: 195 trades taken, 59.7% winners, +0.74R on average, +412.8% total return, average exposure of 99.6% and a peak of 120.1%. It also used a proprietary quality and growth score to choose which signals to take: 431 triggers were skipped.

This is a description of a claim, not a criticism of it. A single window of 4.8 years, with settings tuned live on screen, is a normal way for a backtest to be presented. The question here is a narrower one: does the breakout week itself add anything?

How we tested it

The rules were written down and hashed before any return was computed (preregistration sha256 prefix d0c4964c; the relative-strength follow-up is ae94d9aa). Verdict words were fixed in advance as PASS or INVESTIGATE only.

• Measure: R-multiples under the house exit: a 10% stop from the first daily open of the week after the trigger, plus a 21-week moving-average trail armed from that open. Win rate is not the criterion; the median R being negative is how this kind of method normally looks.
• Nested control: a plain "trigger week versus any week" comparison is only a base rate. Control C1 is every week where the same stock is already in a valid box, liquid, above its 20-week EMA, with MACD above its signal line, but without the breakout trigger. Control C2 is the plain close-above-the-box weeks that fail at least one of the three add-ons (4% clearance, close strength, volume). So the only thing that differs is the trigger, or its add-ons.
• Periods: in-sample 2010 to 2021, out-of-sample 2022 to 2026 (the published window). US and Thailand are reported separately and never pooled.
• Statistic: the monthly mean difference, signal minus control, clustered by entry month. PASS needs the signal mean above the control mean with t of at least 2.0 in both halves.

Result 1: the breakout trigger against the same boxes without it

Cells show signal / control. Mean R is the pooled average per trade.

No cell reaches t = 2.0 in either half. The pooled means favour the trigger in all 8 cells (6 of 8 on the month-matched comparison), but the gaps are small: the largest out-of-sample US gap is +0.128R (signal +0.245 against control +0.117), at t = +1.08. All four verdicts are INVESTIGATE.

In Thailand the 2010-2021 sample is a strong-trend period for every group, controls included (the C1 control averages +0.678R in 2010-2021). In 2022-2026 both the trigger group and the control group have negative mean R, and the trigger group's median is -1.00R.

Result 2: adding the relative-strength gate (RS 80 or above)

The follow-up changed one variable: a point-in-time RS percentile (computed inside the study from price history, not from stored ratings) of 80 or above, applied to the trigger and to matched controls C1r and C2r. H4 is a filter split of the trigger itself into RS 80 or above versus below 80, and is not a control.

All six verdicts are INVESTIGATE. The one cell above t = 2.0 is the Thai in-sample RS split (t = +2.70), and it does not repeat out-of-sample (t = +1.07). Against matched controls, the RS 80 triggers are indistinguishable from RS 80 boxes without a trigger in both markets. In this sample the trigger group is only mildly RS-skewed: about half of triggers (50.1% TH, 53.4% US) already had RS of 80 or above.

Replicating the published backtest exactly

This section was revised on 2026-10-30. An earlier, simpler copy (the video's exit applied to every liquid US symbol, without its portfolio rules) gave different numbers; after independent review the test was rebuilt to follow the video's own settings as closely as the public information allows. The figures below replace the earlier ones.

Setup. US common stocks on NYSE, NASDAQ, AMEX and CBOE with a market cap of about $1B or more, weekly bars, 2022-01-01 to 2026-10-14. Entry is a buy-stop just above the breakout close in the following week. Exits follow the video: partial sales at each R target with the stop raised one level behind, no commissions. The portfolio starts at 100,000, sizes each position at 10% of current equity and caps gross exposure at 120%. The video ranks simultaneous signals with a proprietary quality and growth score that cannot be rebuilt, so here the order is random: each figure is the median of 200 random-priority portfolios. Two details the video does not state were run both ways: whether the box is drawn from weekly closes or from weekly highs and lows, and whether the stop triggers on an intraweek low or on a weekly close.

The S&P 500 returned about +75% over the same window. In the closest reading (box from highs and lows, stop on a weekly close) the 5th to 95th percentile range of total return across the 200 portfolios is +96% to +205%.

• Per trade, the closest reading comes near the claim. Signal count (673 against 626), mean R (+0.66 against +0.74) and maximum drawdown (17% against 16%) are close. The win rate (52% against 59.7%) is lower.
• Total return is about a third of the claim. +148% against +413%; even the 95th percentile of the 200 random-priority portfolios (+205%) is about half the published figure.
• Plausible sources of the remaining gap: the proprietary quality and growth score, used both as a filter and to rank signals; fewer trades taken here (102 against 195); market caps approximated from current share counts rather than historical ones; and ADRs, which are missing from the share data used here. None of these can be checked from outside, so the gap is unexplained rather than attributed to any one of them.
• Breakout against the same box without a breakout, under the video's own exits. Comparing breakout weeks with weeks where the same stock sat in the same kind of box and trend but did not break out, the difference is positive in all four readings but small: month-clustered t of at most +1.07. The verdict is INVESTIGATE: this data does not show that the breakout week itself adds an edge over being in a trending box, and it does not show the opposite either.

What this shows

• The add-ons did not separate from being in a trending box. Stocks already sitting in a valid, trend-confirmed box did about as well per trade as the weeks that carried the full breakout trigger. Much of the measured return belongs to the box-plus-uptrend state, not to the breakout week.
• Selection is a large lever. The published run took 195 of 626 triggers. The video does not say how the 431 skips split between its stop-width cap, a full portfolio (where it ranks candidates by a proprietary quality and growth score) and that score as a universe filter. Selection steps like these can move results a long way, and a score that cannot be rebuilt point-in-time cannot be checked from the outside.
• Exposure and compounding are not per-trade edge. Average exposure near 100% with a peak above 120% and compounding through a strong US stretch produce a large total return even when the per-trade average is modest. In the exact replication the closest reading reaches +0.66R per trade against the claimed +0.74R, while total return is +148% against +413%: most of the remaining distance sits in selection and trade count, not in per-trade edge.
• Thailand 2022-2026 was negative R for both groups. That is a market-regime fact, visible in the controls as much as in the trigger group.

Limits

• The US database is survivor-heavy: delisted names are absent or truncated, which lifts every arm. The RS follow-up additionally keeps only US symbols with a bar after 2026-10-01 (5,348 kept, 7,271 dropped). The tilt is shared by signal and controls, so the comparison is relative, but absolute R levels are optimistic.
• Prices are adjusted, which understates older dollar volumes in the liquidity floor. ETFs and OTC names are not separable in the database and were left in.
• The Thai universe includes listed REITs and property and infrastructure funds, because they trade as stocks; they sit in every arm.
• The liquidity floor replaces the source's market-cap filter; the quality and growth score was not used.
• The house exit differs from the published one. The exact replication follows the video's exits and portfolio rules, but its signal priority is random because the quality and growth score cannot be rebuilt; market caps are approximated from current share counts, and ADRs are not in the share data.
• Consecutive trigger weeks in the same box are not independent; month clustering reduces this but does not remove it. Entry at the signal-bar close was not run as a sensitivity for these cells.
• INVESTIGATE means the test did not find a separation under these rules. It is neither proof that the pattern has no edge nor proof that it has one.

Educational research only. These are descriptions of what historical data shows under stated rules, not a recommendation about any security.