A companion to Zero Alpha, the post about spending real money to find out my backtest was lying. This is the catalog of everything I buried to get there.
I tried a lot of indicators. Twenty-five different strategy files, dozens of filters, every classic on the menu. Most of them looked brilliant at some point — that was the problem. An indicator that never works is easy to throw away. The dangerous ones are the ones that work in one window, light up the backtest, and quietly fall apart everywhere else.
Here’s the graveyard, sorted not by indicator but by how each one fooled me. The failure modes turned out to be more useful than the indicators ever were.
The window illusion
A volume filter was my first great hope. Only take trades when volume is in the top 40% of recent bars — the idea being that big moves need participation. On thirty days of data it was gorgeous: +8.9% versus +2% for the unfiltered version, profit factor 2.25. I shipped it.
Then I ran it on six months. The edge evaporated. The filter hadn’t found a real pattern; it had found the specific thirty days I tested it on.
The time-of-day gate did the same thing, louder. Only trade the opening hour and the power hour into the close — and on six months of two-minute bars it returned +103% with a Sharpe near 3. That’s a hedge-fund number. I genuinely thought I’d found it. Then I checked the same rule on one-minute bars: deeply negative. I checked which days made the money: the top five days produced more profit than the strategy made all year, and the other thirty-eight trading days combined lost money. It wasn’t a strategy. It was five good days with a story wrapped around them.
The lesson I kept relearning: a short backtest on a trending asset doesn’t measure your edge. It measures the window.
The friction illusion
VWAP reversion. Opening-range breakout. Long-short on the same signal. Each one was positive — sometimes wildly positive — with zero trading costs. Each one went underwater the moment I added realistic commission and slippage. VWAP reversion went from healthy to −81%. The breakout went to −32%. The long-short variant flipped from +149% to −18.8% on the identical trades.
These weren’t bad ideas, exactly. They were ideas whose entire profit lived in the gap between “free trading” and “real trading.” At the one-minute scale that gap is everything, and no amount of indicator cleverness closes it.
The high-win-rate trap
This one is counterintuitive enough that it took me a while to believe.
A mean-reversion rule — buy when RSI drops below 30 — won 60% of its trades. Sixty percent! And lost 77% of its money. Selling when RSI ran above 75 was the same poison in reverse: switching it on raised the win rate from 41% to 56% and made the returns worse.
The math is simple once you see it. A high win rate usually means you’re taking profits early and letting losers run to your stop. You win often and small, lose rarely and big. The win rate is the most emotionally satisfying number on the dashboard and one of the most misleading. I learned to distrust any change that made the win rate prettier.
The concentration illusion
On a leveraged ETF called KORU I built a “smart dip entry” that returned +64% and beat buy-and-hold by 15 points. Beautiful — until I broke down where the return came from. Three trades produced 42 of those 64 points. Three. Optimize anything further on that window and I’d just be fitting to three lucky days that will never repeat.
This is the same disease as the window illusion, viewed through a microscope. It’s not enough for the total to look good. You have to ask how many days, how many trades, are actually carrying it — and whether you’d have had the nerve to be in those exact trades.
The most beautiful backtest I ever made was a costume
The crown jewel of the graveyard is a strategy I called Adaptive Leverage Hold. It scaled exposure up in strong trends and down in weak ones, and the backtest was the best number I’ve ever personally generated: +277% versus +97% for buy-and-hold, a Sharpe ratio of 8.39, and it held up across eleven different parameter perturbations. By every metric on my own dashboard, I’d found something special.
I almost didn’t look closely, because who interrogates their best result? When I finally did, the trick fell out. The strategy never actually exited — it just dialed leverage between 1x and 2x on an asset that, over the test window, went almost straight up. I hadn’t built a strategy. I’d built leveraged buy-and-hold wearing a strategy’s costume, backtested on the one regime where leveraged buy-and-hold is unbeatable. The Sharpe of 8.39 wasn’t skill. It was a bull market with the volatility temporarily turned off.
Now, when a result looks too good, my first question isn’t “how do I deploy this” — it’s “what is this secretly long?”
The exits that did nothing
Not every failure was dramatic. Some indicators just quietly didn’t matter.
I added a fixed trailing stop with real conviction. Over a hundred exits, it fired seven times and added a rounding error. An ATR-adaptive trail got overwritten every time by the regime-break exit before it could act. A breakeven stop never armed in live trading. A time-based stop never once met its conditions before something else closed the trade. A scale-out ladder that sold into strength actively hurt — it underperformed a plain trailing stop by 34% because it kept cutting winners to lock in small gains.
These are the quiet graves. No catastrophe, just code I was proud of that turned out to be decoration. They taught me to test whether a feature does anything before believing it does something good.
The regime models
I tried to get clever with a hidden Markov model — let an algorithm discover the market’s hidden “states” instead of hand-coding rules. In the training data it was stunning: clean regimes, 98–99% persistence, exactly the kind of structure you dream of. On new data, the most bullish state’s next-move return flipped negative. The regimes were real patterns in the past and pure noise about the future. Beautiful in-sample fit is not evidence. Sometimes it’s the opposite.
The survivors — and why none of them is a day-trading strategy
A few things lived. None of them is what I set out to build.
The volatility tools work: a big-move radar, a freefall detector, a regime gauge that flagged elevated risk three days before a major crash. But every one of them predicts magnitude, not direction — they tell you when something is about to happen, never which way. Useful for sizing and risk. Useless for picking a side.
The big-move radar in the wild. It will happily tell you a large move is brewing and that volatility, RSI and order flow all agree — then put the direction at a flat fifty-fifty. Magnitude, never direction.
And one actual strategy survived honestly: a slow trend rule on the unleveraged sector fund, +983% over five and a half years against +516% for buy-and-hold, at half the drawdown. It works because it sidesteps the crashes that take a 3x fund 900% to climb out of. It also trades about seven times a year, which is why — to my own continuing irritation — it’s the one validated edge in this entire graveyard that I keep choosing not to run.
That’s the whole catalog. Two dozen indicators, and the through-line is that almost none of them were wrong. They were real patterns, measured correctly, that simply didn’t survive contact with a different window, a realistic cost, or an honest question about where the money actually came from. Learning to ask those questions was worth more than any indicator I buried to find them.