Stop guessing. Start building.

8,370 days of SPY, every one labelled by its ADX reading. A high ADX doesn't point up. Both touch rates rose. Under ADX 15: 78% later touched +1%, 70% touched −1%. At 40 and above: 88% and 73%. The up rate just rose more. https://statoasis.com/overfit/research/how-to-use-the-adx-indicator-like-a-pro-step-by-step-guide

Since 1929, not one calendar year escaped without a drawdown. The calmest year ever (1995) still dipped −2.5%. The median year fell 13.2% below its own high at some point. And yet 67% of years finished positive anyway — 73% with dividends. Of the 60 years with a double-digit dip, 48% still ended the year up. The dip isn't a warning. It's the toll. Every S&P 500 drawdown since 1871, measured: https://statoasis.com/overfit/research/sp500-drawdowns-since-1870

The 2007 crash fell −57% — nearly as deep as anything since 1929. Investors were back at new highs in 5.5 years. The 1881 decline fell "only" −42%. It kept investors underwater for 19.5 years. Depth is set by the panic. Duration is set by what the economy does after. Every S&P 500 drawdown since 1871, measured: https://statoasis.com/overfit/research/sp500-drawdowns-since-1870

The "~90% win rate RSI" is the most-shared claim in retail trading. It doesn't survive a rebuild. The claim comes from a Triple-RSI mean-reversion rule: RSI falling three bars, oversold, price above its long trend. Rebuilt on SPY from 1993, the RSI(2) version wins 59.1% of trades at a 5-bar hold and 65.1% at a 10-bar hold. The canonical RSI(5) version lands at 58.9% and 64.5%. Positive, profitable, and nowhere near 90%. The gap most likely traces to a volatility filter in the original ("RSI below its value 63 bars ago") that this clean rebuild leaves out. The full rebuild across SPY, QQQ, IWM, and DIA: https://statoasis.com/overfit/research/better-rsi-backtest

If you run a MACD, here is the change, in four lines. Add an OBV MACD as a second panel. Take a bullish price crossover only when the OBV MACD histogram is already above zero. Mirror it for shorts. Leave your exit exactly as it is. That is the whole edit. It took 9,216 backtests on six markets to justify it and about thirty seconds to implement. What to expect, so you are not surprised into abandoning it: You will trade about 42% less. Six signals a year instead of ten at the default setting. In every one of 3,072 matched pairs, the filtered version took fewer trades than its control. There were no exceptions, which is what a filter is supposed to look like. Your net profit will probably go down on the long side. It improved in only 30.4% of pairs there. Skipping trades in a thirty-year uptrend means skipping winners. Your drawdown will probably go down more. It fell in 85.7% of pairs, by a median of 21.3 percentage points. One setting note, because it decides whether this works. Use the slow end: fast 12 or more, slow 35 or more, signal 14 or more. That corner was profitable on all six markets in 91.7% of its 48 configurations, at a median 1.64 profit factor. The fast corner managed 53.1% and 1.26. Forty-eight configurations, not one lucky cell. That is what makes it worth trading rather than worth screenshotting. And a floor under the whole thing: none of these crossovers beat holding the index. What they buy is a shallower drawdown. Know which one you are buying. The parameter maps, the six-market breakdown and the limits: https://statoasis.com/overfit/research/obv-macd-vs-traditional-macd-which-one-wins

1,856 backtests across four ETFs, and "better RSI" turns out to be barely better. Plain RSI, Connors RSI, Laguerre RSI, and Z-Score RSI were run over SPY, QQQ, IWM, and DIA — every length, threshold, side, and time-exit. The question was simple: do the upgrades actually beat the original? Z-Score RSI posts the best median annual return of the four at +1.05%, versus plain RSI's -0.15%. Connors RSI also edges ahead at +0.32%. Laguerre RSI does the opposite — it trails the baseline at -0.66%. Two upgrades win, one loses, and the winning margin is measured in fractions of a percent. Full family-by-family tables and the best variant in each: https://statoasis.com/overfit/research/better-rsi-backtest