Stop guessing. Start building.

33,792 backtests of RSI vs Stochastic vs Williams %R: the three tie as mean-reversion longs — 100.0%, 99.8%, 89.6% of variants profitable. The real split is the setup. Long variants 89.7% profitable, short 8.8%. Full three-way study: https://statoasis.com/overfit/research/rsi-vs-stochastic-vs-williams-r

I ran 33,792 backtests to settle it: RSI vs Stochastic vs Williams %R, on four US index ETFs. The question "which oscillator is best?" turned out to be the wrong one. Here's what actually decides it. 🧵

I tested the three ADX rules everyone repeats, on 8,370 SPY days and 4,889 E-mini days. Two held up. The one about direction didn't. The strongest region out of 3,780 variants was long inside a "downtrend" reading. Everything you need is here: https://statoasis.com/overfit/research/how-to-use-the-adx-indicator-like-a-pro-step-by-step-guide

"How far can it fall" is the question people ask. "How long until it comes back" is the one you actually live with. We measured every S&P 500 drawdown of 10% or more since 1871, 32 episodes in all, and timed each from its low back to a new high. Sort them by depth and the waiting time sorts itself. - Falls of 10% to 20%, 15 episodes: median 94 calendar days back to a new high. - 20% to 30%, 6 episodes: 386 days. - 30% to 40%, 5 episodes: 639 days. - 40% or deeper, 6 episodes: 1,638 days. Pool every episode 15% or deeper and split it by era. The 6 that finished before 1950 took a median 1,112 days to get back. The 15 since took 337. Those are two pooled medians. They say recoveries since 1950 have been faster across all depths together, not which rung of the ladder moved. And the deep end is thin. Six episodes is six episodes. If you are sizing around the worst case, read 1,638 days as a sense of scale, not a forecast. Every drawdown since 1871, depth and recovery time side by side: https://statoasis.com/overfit/research/sp500-drawdowns-since-1870

33,792 backtests later, "which oscillator is best?" turned out to be the wrong question. I ran RSI, Stochastic and Williams %R against each other on four US index ETFs — SPY, QQQ, IWM and DIA — across mean-reversion and breakout, long and short, every combination. Then I asked the same question everyone asks: which indicator wins? They tie. As a mean-reversion long, 100.0% of Williams %R variants were profitable, 99.8% of Stochastic, and 89.6% of RSI. Median R-expectancy across the three: 0.20, 0.20, 0.18. Three different formulas, one edge. What actually separated winners from losers wasn't the oscillator. It was the setup: 89.7% of long variants were profitable versus 8.8% of short ones. The indicator is a detail. The setup is the decision. Full three-way breakdown: https://statoasis.com/overfit/research/rsi-vs-stochastic-vs-williams-r

The win-rate statistic failed in both directions in the same study — and each failure is a lesson. We ran 1,856 RSI-family backtests across SPY, QQQ, IWM, and DIA. Sort the results by win rate and look at both ends of the leaderboard: The top: a plain-RSI configuration on QQQ with a 100% win rate. Perfect record. It took 5 trades in 27 years and made $1,900. Every single one of the ten highest win rates in the grid sits below 50 trades — which is exactly why the engine flags anything under that threshold as unreliable instead of celebrating it. The middle: a short-side configuration on IWM with a 60.6% win rate across 582 trades — a genuinely large sample — that lost $4,592. Wins six trades in ten and still loses money, because the average loser is bigger than the average winner. So a win rate can lie two ways: a perfect rate on a sample too thin to mean anything, and a good rate whose losers quietly outweigh its winners. The two numbers that keep it honest are the trade count and the win/loss size ratio — and neither ever appears in a screenshot. Full study — four RSI families, 1,856 variants, every number traceable: https://statoasis.com/overfit/research/better-rsi-backtest