> ## Content Index
> Fetch the complete content index at: https://simplestocktrade.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Three lower lows: a real edge that faded into a coin flip
- URL: https://simplestocktrade.com/three-lower-lows-edge-decay/
- Published: 2026-11-17T16:06:56.000Z
- Updated: 2026-11-17T16:40:04.000Z
- Description: A classic mean-reversion rule, run the way an account would trade it on 29 years of large US stocks: 9% a year until 2006, about 1% since 2016, and a coin flip now does as well. Daily and weekly bars.
- Author: Yevhen S
- Tags: Mean reversion, Strategy reviews, backtesting

**The short version.**

- **What it does:** buys a large US stock after three lower lows in a row while it is still in an uptrend, with a limit order half an ATR under the close, and sells at the next open after an up close.
- **What it returned:** 4.6% a year from 1998 to November 2026, worst drawdown −24.7%, in the market a third of the time. The S&P 500 with dividends did 9.1% a year over the same stretch.
- **The finding:** the edge was real and it faded. 9.1% a year in 1998–2006, 4.6% in 2007–2015, 1.1% since 2016.
- **The test that decides it:** replace the pullback signal with a coin flip and keep everything else. Before 2007 the real signal beat every one of 100 coins. Since 2016 it beats 55% of them, which is what a coin does.
- **Weekly bars do not rescue it:** 4.1% a year with a −39% drawdown, and the signal never clearly beat a coin there either. One deeper weekly variant looks alive, but it was found after the fact (last section).
- **Verdict:** a textbook mean-reversion idea that worked, got crowded, and is now the market with extra steps. Backtest of my own; not advice.

## What this is

A short-term mean-reversion strategy on large US companies. The idea is old and sound: in a stock that is trending up, a few days of selling are usually noise, and buying them pays when the stock snaps back. This version measures “a few days of selling” as three lows in a row, each under the one before.

I ran it the way an account would actually trade it, on 29 years of data, and then tried to break it. It did not break in the way I expected. It did not stop working on one bad day or in one crash. It faded, era by era, until the signal stopped carrying any information at all.

## The rules, exactly

- **Universe:** stocks in the Russell 1000 on the day, including companies that later left the index or stopped trading.
- **Setup, at the close:** close above its 100-day average, close under its 5-day average, and three lower lows in a row (each day’s low under the previous day’s low).
- **Filters:** 21-day average dollar volume above $10 million, price above $1\. I added one more: the S&P 500 must be above its own 200-day average, otherwise no new orders.
- **Entry:** a limit order for the next day at the setup close minus 0.5 × the 10-day ATR. It fills only if the price trades through the limit. A touch is not enough.
- **Exit:** sell at the next open after any close above the previous close. A 3% profit target is in the file too, but it almost never acts first (see below).
- **Book:** ten positions of 10% of the account, no leverage, $0.01 a share in commissions. When there are more setups than free places, the order list is drawn at random, as in the original test.

Two details decide whether a backtest of this kind can be traded, and published versions often get them wrong:

1. **The place in the book goes to the order, not to the fill.** A limit order that never fills still took the place it was sent for. A backtest that picks its ten names from the orders that already filled is choosing with hindsight.
2. **The open exit comes before the target.** When the rules say “sell at the open”, the order goes out before the open, and the account cannot also collect the target on the same day. Getting this wrong was worth half the return on this strategy.

## Where the orders actually go

2026-11-17T15:30:24.200211 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 

Pfizer, April 2005: the third lower low, the limit half an ATR under the close, the fill, and the sale at the open

Pfizer in April 2005\. On 27 April the stock closed above its 100-day average, under its 5-day average, and printed its third lower low in a row. The order for the next morning: buy at 25.27, half an ATR under that close. On 28 April the price traded down to 25.20 and the order filled at 25.27\. On 29 April the stock closed higher than the day before, so it was sold at the open on 2 May, at 25.82\. That is +2.2% before costs, in three trading days, and it is what this strategy looked like when it worked.

## I ran it

3 lower lows, 1998-2026 Strategy stats - a backtest, not a forecast Strategy S&P 500 TR Period 1998-01-01 .. 2026-11-16 1998-01-01 .. 2026-11-16 Annual return (CAGR) 4.6% 9.1% Max drawdown \-24.7% \-55.2% Return / drawdown (MAR) 0.2 0.2 Sharpe (daily, no risk-free) 0.5 0.6 Sortino 0.6 0.7 Volatility (annualised) 9.4% 19.3% Best / worst month +10.1% / -9.7% +12.7% / -16.5% Positive months 50% 63% Longest time under water 1409 days 2404 days Time in the market 34% 100% Trades (per year) 9596 (332) Win rate 60% Profit factor 1.1 Average trade +0.2% Average win / loss +1.9% / -2.5% Average holding period 4 calendar days Russell 1000 as it stood each day - limit orders, slot spent on the order - open exits before targets - 0 USD a share - S&P 500 TR = the S&P 500 ETF with dividends 

Strategy statistics 1998-2026 next to the S&P 500 with dividends

3 lower lows, 1998-2026 Combined monthly percent gains - green up, red down; the sign is printed too Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Year S&P 500 TR 1998 \-0.2 +4.0 \-3.7 +4.7 +0.7 +0.9 \-4.8 \-4.8 +0.5 +0.1 +3.7 \-1.5 \-0.9 +28.0 1999 +2.6 +4.2 +2.4 +7.0 +4.2 \-2.4 +2.9 +2.5 \-1.7 \-2.0 +5.9 +3.6 +33.0 +20.4 2000 \-2.1 \-2.0 +9.2 +4.5 +4.7 +1.5 +0.1 +5.8 \-3.8 \-0.9 0.0 0.0 +17.6 \-9.7 2001 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 \-11.8 2002 +0.1 0.0 +2.9 +0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 +3.2 \-21.6 2003 0.0 0.0 0.0 +0.9 +4.3 +6.1 +0.8 \-0.4 \-0.4 +0.2 +0.4 \-1.7 +10.3 +28.2 2004 +1.6 +2.5 \-1.7 \-1.2 +2.2 +1.0 \-1.5 0.0 +0.4 \-1.0 +2.8 +3.6 +8.7 +10.7 2005 \-1.6 +1.1 \-0.7 \-1.5 +0.5 +3.8 +1.4 \-1.3 \-2.5 \-2.1 +3.2 +1.7 +1.7 +4.8 2006 +5.1 \-0.8 +2.1 +2.4 \-2.8 \-0.1 \-4.2 +1.8 \-0.5 +2.9 +3.2 +3.3 +12.7 +15.8 2007 \-0.9 \-1.7 \-2.7 +2.5 +0.7 \-1.9 \-0.1 +3.0 +1.4 \-2.3 \-2.7 +1.0 \-3.8 +5.1 2008 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 \-36.8 2009 0.0 0.0 0.0 0.0 0.0 \-1.2 +1.6 +1.6 +2.4 +0.2 +0.9 +1.8 +7.6 +26.4 2010 \-3.4 +1.2 +4.2 +4.9 \-2.4 \-1.3 0.0 +0.2 +3.5 +2.3 +0.2 +2.4 +12.3 +15.1 2011 +2.0 +2.1 +2.6 +0.4 +0.3 \-0.8 \-5.7 \-3.0 0.0 \-0.2 \-1.8 \-0.3 \-4.7 +1.9 2012 +1.5 +5.0 +2.0 +2.4 \-4.0 \-0.6 \-1.9 +3.6 +1.4 +4.0 +0.8 +2.5 +17.5 +16.0 2013 +2.3 +1.0 +2.0 \-1.9 +0.7 +0.6 +3.2 \-2.5 +1.5 +0.8 +2.1 +2.5 +12.9 +32.3 2014 +0.8 +2.3 +3.1 0.0 +2.3 +2.7 \-4.3 +2.5 \-3.4 \-4.0 +1.7 \-0.7 +2.5 +13.5 2015 \-0.6 +3.2 +0.2 +1.7 \-0.6 +1.2 \-1.3 \-3.8 0.0 +0.1 \-0.5 \-0.7 \-1.3 +1.2 2016 \-0.5 0.0 +1.0 \-3.9 +1.6 +0.2 +0.6 +1.7 \-0.1 +0.1 +4.8 \-0.1 +5.3 +12.0 2017 +0.1 +0.2 \-3.3 \-1.2 \-1.9 +1.6 \-0.5 +1.4 +1.7 \-0.6 +4.7 0.0 +2.0 +21.7 2018 +0.5 \-4.3 \-1.0 +0.5 +3.1 0.0 +1.8 +0.6 +0.6 \-3.5 +0.1 0.0 \-1.8 \-4.6 2019 0.0 +1.5 \-1.3 \-1.9 \-6.8 +1.1 +0.9 \-3.4 +1.6 \-1.9 +4.3 +1.5 \-4.9 +31.2 2020 \-4.2 \-9.7 +0.9 0.0 +0.6 \-4.4 +6.0 +2.7 \-2.9 +4.1 +4.5 +2.4 \-1.2 +18.3 2021 \-0.4 +3.3 +2.4 +2.6 +0.1 \-2.5 +7.1 \-0.7 \-1.9 +0.7 +1.0 +4.1 +16.5 +28.7 2022 \-4.6 0.0 +3.3 \-2.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 \-0.7 \-4.6 \-18.2 2023 \-1.2 \-0.1 \-3.4 +0.1 \-2.4 +1.1 +1.2 \-1.5 \-3.2 \-0.4 +1.7 +3.1 \-5.0 +26.2 2024 \-1.6 +1.6 +1.9 +0.6 \-1.9 \-0.2 \-0.6 +2.6 \-0.8 \-0.5 +1.2 \-5.5 \-3.6 +24.9 2025 +7.3 \-1.6 \-1.5 0.0 +0.3 +3.5 +0.3 +2.3 +1.8 \-3.4 +8.0 \-6.4 +10.0 +17.7 2026 +1.6 +2.3 \-4.2 \-1.3 \-3.6 +1.9 +10.1 +2.5 \-3.6 +0.9 \-3.8 +2.1 +9.4 Russell 1000 as it stood each day - limit orders, slot spent on the order - open exits before targets - 0 USD a share - S&P 500 TR = the S&P 500 ETF with dividends 

Month by month, 1998 to 2026, with each year and the S&P 500 with dividends

2026-11-17T16:25:36.8 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 100 1,000 1998-2006 +9.1% a year 2007-2015 +4.6% a year 2016-2026 +1.1% a year 3 lower lows (4.6% a year) S&P 500 with dividends (9.1% a year) 2000 2004 2008 2012 2016 2020 2024 2028 −40 −20 0 below the last high, % Twenty-nine years of 3 lower lows Growth of 100 on a log scale, against the S&P 500 with dividends. Bottom: how far each sat below its own high. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% 

Growth of 100 on a log scale against the S&P 500 with dividends, drawdowns underneath

Russell 1000 as it stood each day, January 1998 to November 2026, ten positions of 10%, $0.01 a share, orders placed before the open. **4.6% a year with a −24.7% worst drawdown, in the market 34% of the time.** That beats the S&P 500 on drawdown and loses to it badly on return. The flat stretches on the curve are late 2000 to early 2003, 2008 to mid-2009 and most of 2022: the index filter kept the book out of the bear markets, which is why the drawdown is shallow.

The whole-period number hides the story. Split into three eras, it is three different strategies:

2026-11-17T16:25:36.9 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 1998-2006 2007-2015 2016-2026 0% 4% 8% 12% 16% annual return 9.1% worst -11% 5.8% worst -48% 4.6% worst -12% 6.3% worst -55% 1.1% worst -25% 14.4% worst -34% 3 lower lows S&P 500 with dividends Three eras, three different strategies Annual return and the deepest drawdown of each era. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% 

Annual return and deepest drawdown in 1998-2006, 2007-2015 and 2016-2026, against the index

2026-11-17T16:25:36.9 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 2000 2005 2010 2015 2020 2025 \-0.6% \-0.4% \-0.2% +0.0% +0.2% +0.4% +0.6% +0.8% +1.0% average trade, by year of entry 1998-2006 +0.3% a trade 2007-2015 +0.2% a trade 2016-2026 +0% a trade The edge per trade, year by year Bars: average result of a trade entered that year, after costs. Blue lines: the average of each era. No bar = the index filter kept the book out (2001, 2008). Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% 

Average trade by year of entry, with the average of each era

The average trade, after costs, went from +0.31% in 1998–2006 to +0.15% in 2007–2015 and +0.04% since 2016\. The win rate barely moved (62%, 60%, 59%). What disappeared is the size of the snap-back.

## What I tried to break

| Test                                                                       | What it asks                                            | Daily bars                                                         | Weekly bars                                                                                  |
| -------------------------------------------------------------------------- | ------------------------------------------------------- | ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------- |
| 75 versions of the rules, chosen on 1998–2015                              | Is there a broad profitable zone, or one lucky setting? | 75 of 75 profitable, median 6.2% a year                            | 75 of 75 profitable, median 3.6%                                                             |
| The same 75 on 2016–2026                                                   | Does the zone survive data it was not chosen on?        | 65 of 75 profitable, median 1.2%                                   | 53 of 75 profitable, median 0.9%                                                             |
| 1998–2015 split in half                                                    | Was it already fading before the hold-out?              | Median 8.7% → 3.9%                                                 | Median 5.0% → 1.9%                                                                           |
| Coin-flip entries (100) against the real rule (20 draws of the order list) | Does the pullback signal carry information?             | Beats 100% of coins in 1998–2006, 70% in 2007–2015, 55% since 2016 | Beats 88% of coins in 1998–2006, 13% in 2007–2015, 74% since 2016; 67% over the whole record |
| Extra slippage, 1 to 10 basis points per side                              | How much cost can it take?                              | −0.7 points a year per basis point; since 2016 negative at 3       | −0.15 points per basis point; since 2016 still +1.4% at 10                                   |
| A touch of the limit counts as a fill                                      | Do the fill rules flatter it?                           | 4.64% against 4.63%                                                | 4.14% against 4.14%, the same 2,183 trades                                                   |
| 20 draws of the order list                                                 | How much of one backtest is luck?                       | 2016–2026: −0.7% to 4.6% a year                                    | Whole record: 2.5% to 7.8% a year; 2016–2026: 3.0% to 8.3%                                   |

**The parameters were not the problem.** I ran 75 versions of the rules: two, three or four lower lows, the limit from 0.25 to 1.25 ATRs under the close, the target off or from 1.5% to 6%. The grid was read on 1998–2015 only, and the choice written down before the same grid ran on 2016–2026\. (One caveat: the published settings themselves I had already run over the whole period once, in September, so for that single cell the last decade is not a blind test. For the other 74 it is.) On 1998–2015 all 75 made money, median 6.2% a year. On 2016–2026, 65 of 75 still made money, but the median was 1.2%. Choosing well did not help: the fifteen best versions of the first period made 1.8% a year in the second, against 1.6% for all of them.

2026-11-17T16:25:36.961379 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 2 3 4 lower lows in a row 0.25 0.50 0.75 1.00 1.25 limit, ATRs under the close 5.3 6.9 4.9 7.9 6.8 6.0 7.5 7.4 6.5 6.8 5.9 5.5 4.8 4.8 4.3 1998-2015 (where the settings were chosen) 2 3 4 lower lows in a row 0.25 0.50 0.75 1.00 1.25 limit, ATRs under the close 0.4 0.3 3.5 1.8 1.2 4.3 2.6 \-0.6 4.1 0.9 0.4 2.7 \-0.1 1.2 1.7 2016-2026 (never used to choose) Every setting faded at the same time Annual return % for 15 versions of the rules (3% target). The rules as published: 3 lows, 0.5 ATR. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0.01 a share · 10 positions of 10% 

Fifteen versions of the rules: 1998-2015 against 2016-2026

**The coin flip.** This is the test that matters. I kept everything about the strategy — the universe, the trend and index filters, the limit half an ATR down, the exits, the costs, the number of orders — and replaced only the pullback signal with a coin flip. A hundred coins, against the real rule run on twenty different random draws of the order list.

2026-11-17T16:38:32.6 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 0% 4% 8% 12% 0 2 4 6 8 10 12 random-entry runs 1998-2006 the rule beats 100% of coins 0% 2% 4% 6% 8% 0 2 4 6 8 10 12 2007-2015 the rule beats 70% of coins \-2% 0% 2% 4% 6% 0 2 4 6 8 10 12 14 2016-2026 the rule beats 55% of coins 2% 4% 6% 0 2 4 6 8 10 12 14 1998-2026 the rule beats 99% of coins Did the signal know anything a coin did not? Grey: the pullback trigger replaced by a coin flip - same filters, limit, exits, costs, matched order count. Blue: the real rule on 20 ranking seeds; the thick line is their median. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% 

The real rule against 100 coin-flip entries, era by era

In 1998–2006 the real rule beat all 100 coins: 9.8% a year against a coin median of 3.7%. In 2007–2015 it beat 70% of them. Since 2016 it beats 55%: the signal is now worth exactly what a coin is worth. What is left of the return comes from what the coin also has — buying an uptrending stock on a limit under the market and selling it on the first up close.

**Costs.** The edge per trade is now a few hundredths of a percent, so it cannot pay for anything. Every basis point of extra slippage on each side costs about 0.7 points a year over the whole record. Since 2016 one basis point leaves 0.4% a year, and three turn it into a loser.

2026-11-17T16:25:37 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 0 2 4 6 8 10 extra slippage per side, basis points (on top of $0 a share) \-6% \-4% \-2% +0% +2% +4% +6% +8% annual return 1998-2006 2007-2015 2016-2026 1998-2026 How much slippage can it take? Annual return when every entry and exit costs a little more than the backtest assumes. Each basis point per side costs the whole record about 0.7 points a year. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% 

Annual return when every entry and exit costs 0 to 10 basis points more

## The same rules on weekly bars

2026-11-17T16:25:24.8 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 1998-2006 2007-2015 2016-2026 0% 4% 8% 12% 16% annual return 9.1% \-11% 4.6% \-12% 1.1% \-25% 7.9% \-12% 1.5% \-18% 3.2% \-39% 5.8% \-48% 6.3% \-55% 14.4% \-34% daily bars weekly bars S&P 500 with dividends Daily and weekly bars, era by era Annual return of each era, with its deepest drawdown underneath. Same rules; the weekly version counts weeks instead of days. Russell 1000 as it stood each day · weekly bars: every length in weeks, limit good for the next week, sold at the next week's open · $0 a share · 10 positions of 10% 

Daily and weekly versions against the S&P 500 with dividends, era by era

The obvious objection to a decay story: markets got faster, so a slower version of the same idea might still work. I ran the rules again with weeks in place of days. A 100-week and a 5-week average, three lower weekly lows, a limit half a 10-week ATR under the week’s close, good for the whole next week. A sale at the next week’s open after an up week. New orders only while the S&P 500 is above its 200-week average.

**1998–2026: 4.1% a year with a −39% worst drawdown, in the market 35% of the time.** 2,183 trades, average +0.61%, 65% winners. By era: 7.9%, 1.5% and 3.2% a year.

**Costs barely touch it.** A weekly trade lasts about three and a half weeks and averages four times the daily one, so ten basis points of extra slippage on each side still leave the last decade at 1.4% a year.

**The drawdown is the price.** Almost all of the −39% came in four weeks. By the end of February 2020 the book was full, ten positions, nine of them bought in the second half of the month as the market turned, and the weekly exit waits for an up week. One cruise-line stock lost 78% before the rules could sell it. The daily version, out after the first up day, lost 12% over the same weeks; the weekly one rode the whole crash. Getting back to the old high took until March 2022.

**The grid tells the same story.** On 1998–2015 all 75 weekly versions made money, median 3.6% a year, and split in half the median went from 5.0% to 1.9%. On 2016–2026, 53 of 75 made money, median 0.9%. (This is not a blind test: I had run the weekly rules over the whole period before the grid, so read it as the shape of the zone, not as a hold-out.) The versions that were best before 2016 made 1.8% a year after it, against 1.4% for all of them.

2026-11-17T16:25:24.816716 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 2 3 4 lower weekly lows in a row 0.25 0.50 0.75 1.00 1.25 limit, ATRs under the close 6.1 4.8 3.8 2.9 4.7 2.7 3.2 3.3 3.3 3.4 3.0 3.7 4.0 2.8 2.0 1998-2015 (where the settings were chosen) 2 3 4 lower weekly lows in a row 0.25 0.50 0.75 1.00 1.25 limit, ATRs under the close 4.5 4.0 2.5 4.2 2.6 0.9 2.8 2.5 \-0.1 0.5 0.2 \-1.5 \-0.5 \-0.2 \-2.2 2016-2026 The weekly grid, before and after 2016 Annual return % for 15 versions of the weekly rules (3% target). The rules as published: 3 lows, 0.5 ATR. Russell 1000 as it stood each day · weekly bars: every length in weeks, limit good for the next week, sold at the next week's open · $0.01 a share · 10 positions of 10% 

The weekly grid: 1998-2015 against 2016-2026

**The coin.** This is where the weekly bars part company with the daily ones. On daily bars the signal once beat every coin; on weekly bars it never clearly did. The real weekly rule beat 88% of 100 coin-flip versions in 1998–2006, 13% in 2007–2015 and 74% since 2016, 67% over the whole record: inside the coins’ range every time. A weekly book makes about 75 trades a year, so a single run is noisy. The same weekly rules returned anywhere from 2.5% to 7.8% a year over the whole record, depending only on the draw of the order list. The coins made a median 3.8% a year with a −29% worst drawdown, the real rule 4.3% with −37.5%. Whatever the weekly version earns, the scaffolding earns it: an uptrending stock bought under the market and sold on the first up week. The three lower lows add nothing a coin does not.

2026-11-17T16:38:34.1 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 0% 3% 6% 9% 12% 0 5 10 15 20 random-entry runs 1998-2006 the rule beats 88% of coins 0% 2% 4% 6% 8% 0 2 4 6 8 10 12 14 2007-2015 the rule beats 13% of coins 0% 3% 6% 0 2 4 6 8 10 12 14 2016-2026 the rule beats 74% of coins 2% 4% 6% 8% 0 2 4 6 8 10 1998-2026 the rule beats 67% of coins Weekly bars: did the signal know anything a coin did not? Grey: the weekly pullback trigger replaced by a coin flip - same filters, limit, exits, costs, matched order count. Purple: the real weekly rule on 20 ranking seeds; the thick line is their median. Russell 1000 as it stood each day · weekly bars: every length in weeks, limit good for the next week, sold at the next week's open · $0 a share · 10 positions of 10% 

Weekly bars: the real rule against 100 coin-flip entries

## What survives

**Buying pullbacks in an uptrend still beats doing nothing with the cash.** The coins themselves made a median 3.0% a year with a median worst drawdown of −26%, against the real rule’s −25%. That is the exit rule and the trend filter, not the signal.

**The index filter earns its place.** It kept the book out of late 2000 to early 2003 and of 2008 to mid-2009, which is why the worst drawdown is −25% against −55% for the index.

**The execution assumptions were not flattering it.** Letting a touch of the limit count as a fill changed nothing (4.64% against 4.63%). The honest order handling had already been applied.

## What breaks

**The signal itself.** It beat every coin before 2007 and no more than half of them since 2016\. Nothing in the rules changed; the market did.

**Any extra cost.** At today’s edge per trade, three basis points of slippage on each side turn the last decade negative.

**Reading one backtest as the result.** The order list is drawn at random. On 2016–2026 the same rules returned anywhere from −0.7% to 4.6% a year depending only on that draw. A single run of this strategy tells you very little.

**The slower clock.** On weekly bars costs stop mattering, but the signal never separated from a coin, and a limit order that stays open all week fills straight into a crash: −37% in four weeks of 2020.

**Copying the published numbers.** The original test reported 22% a year for 2004–2014, in the market 70% of the time. I could not rebuild it, and I will not guess at the reasons. What I can measure is one shortcut: the same rules, allowed to collect the 3% target on mornings they were supposed to sell at the open, made 9.9% a year over 1998–2026\. With the open exit first, as an account would actually trade, 4.9%. Half the return was the shortcut.

## Verdict

This was a real edge, and the kind of thing worth knowing about: a simple, logical rule that made 9% a year with an 11% drawdown for nine years. It is now gone. The decline was gradual, it showed up first in the size of the average trade, and a coin flip with the same exits now does as well. The weekly version of the published rule does not rescue it: costs matter less there, but the signal was never clearly better than a coin and the drawdown is deeper. I would not trade it as it stands. The part worth keeping is the method: when a strategy fades, the random-entry control tells you whether the signal or the scaffolding was earning the money.

## Can you optimise it back to life?

The obvious next move with a fading strategy is to tune it until it works again. After the article went up I tried the simplest knob, the number of lower lows in a row, from two to eight, on daily and weekly bars, with everything else unchanged.

2026-11-17T17:27:17.7 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 2 3 4 5 6 7 8 lower lows in a row 0% 2% 4% annual return 5.5 4.6 5.4 4.5 3.1 1.7 0.2 Daily bars 1998-2026 2016-2026 2 3 4 5 6 7 8 lower lows in a row 3.4 4.1 1.9 5.1 3.2 0.9 0.5 Weekly bars 1998-2026 2016-2026 Seven versions of the rule, one at a time Annual return for 2 to 8 lower lows in a row, everything else unchanged. Dark: 1998-2026\. Light: since 2016. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% · 1998-2026 

Annual return for 2 to 8 lower lows in a row on daily and weekly bars, 1998-2026 and since 2016

On daily bars the last decade is ragged: from −0.6% to 4.3% a year depending on the setting, with neighbours jumping around (three lows 1.1%, four 4.3%, five 2.6%). On daily bars one draw of the order list alone moves the last decade by five points, so none of that is information.

On weekly bars one version stands out. **Five lower weeks in a row: 5.1% a year over 1998–2026 with a −22% worst drawdown, in the market 19% of the time, and no fade at all: 5.7%, 4.1% and 5.4% a year by era.**

It passes the tests the published rule failed. Against 100 coin flips with the same number of orders it beats 100% of them in 1998–2006, 79% in 2007–2015 and 99% since 2016\. Twenty draws of the order list give 4.5% to 6.1% a year, so it is not one lucky draw. The average trade is +1.2%, so costs hardly matter, and letting a touch of the limit count as a fill changes nothing.

2026-11-17T17:27:19.4 image/svg+xml Matplotlib v3.1.8, https://matplotlib.org/ 0% 3% 6% 9% 0 2 4 6 8 10 random-entry runs 1998-2006 the rule beats 100% of coins 0% 2% 4% 6% 0 2 4 6 8 10 12 14 2007-2015 the rule beats 79% of coins \-2% 0% 2% 4% 6% 0 2 4 6 8 10 12 2016-2026 the rule beats 99% of coins 0% 2% 4% 6% 0 2 4 6 8 10 12 1998-2026 the rule beats 100% of coins Weekly bars, 5 lower lows: signal or luck? Grey: the trigger replaced by a coin flip - same filters, limit, exits, costs, matched order count. Purple: the real rule on 20 draws of the order list; the thick line is their median. Russell 1000 as it stood each day · limit orders placed before the open, slot spent on the order · open exits before targets · $0 a share · 10 positions of 10% · 1998-2026 

Weekly bars, 5 lower lows: the real rule against 100 coin-flip entries, era by era

| Weekly bars (median of 20 draws; coin: of 100)                 | 1998–2026 | Since 2016 | Worst drawdown | In the market |
| -------------------------------------------------------------- | --------- | ---------- | -------------- | ------------- |
| 4 lower lows                                                   | 3.1%      | 1.9%       | −38%           | 30%           |
| 5 lower lows                                                   | 5.1%      | 5.3%       | −22%           | 19%           |
| 6 lower lows                                                   | 3.2%      | 3.6%       | −9%            | 10%           |
| Coin flip, same order count as 5 lows                          | 2.0%      | 1.9%       | −21%           | 21%           |
| 5 lower lows on small caps outside the Russell 1000 (one draw) | 3.0%      | 4.3%       | −22%           | 19%           |

So why is this not the happy ending?

**It was found by looking.** Five lower lows was not in the grid fixed before the tests. It turned up after the whole history had been seen, as the best of seven settings on two bar lengths. The best of fourteen will always look better on the past than it can be on the future, and there is no untouched period left to check it on.

**It is a peak, not a plateau.** Its neighbours, over the same 20 draws each: four lower weeks 3.1% a year with a −38% drawdown, six lower weeks 3.2% with −9%. A real effect rarely jumps from 3% to 5% and back between neighbouring settings.

**It does not travel.** On smaller companies outside the Russell 1000 the same weekly rules make 3.0% a year with five lows, and five is nothing special there: three lows make 4.0%, six 1.8%.

**What would make it interesting** is data nobody has seen yet: write the rules down today and let them run untouched, on paper, for a year or two. If the deeper weekly pullback keeps beating the coin there, it becomes a new strategy worth testing properly. Until then it is a backtest that was allowed to choose its own answer.

💡

****Good strategies are rare.** Between September and November 2026 I took about thirty strategy ideas through tests like these, from books, blogs, videos and my own notes. Most ended at “no edge once the orders are placed honestly” or “the index with extra steps”. A few had an edge too thin to survive costs. This one is in the minority that had a real edge at all, and it still faded. A strategy that survives every test above, after costs, on data it was not built on, is the exception. If a backtest looks easy, the first suspect is the backtest.

*This is a backtest of rules I wrote down and ran myself, on historical data, under the conditions stated. It describes the past and is not a forecast or a recommendation.*