The best and worst trading indicators, ranked across 222,010 backtests. The top 4 are not what you expect.

Where this goes next

What a corpus this size is actually for

A ranking is a survey. Nobody was randomly assigned an indicator, every strategy here was written by a person who chose their own tickers and rules, and people reach for different tools on different problems. The ordering above tells you where to look, not what causes what.

222,010 outcomes is enough to stop guessing and start testing. Three things become possible here that a smaller dataset cannot support.

1. Run the experiment instead of the survey

Take one strategy, one universe, one window, one rebalance cadence, and change exactly one thing. The indicator goes in, or it comes out. Everything else is held. Run both.

That is a controlled experiment, and it is a backtest sweep the platform already knows how to run. The ranking's job is telling you where to point it. You cannot sweep everything, and 31 candidates ordered by median Sharpe is a very good prior.

2. Go down to the configuration, not the type

An indicator type is not one rule, and treating it as one rule is the weakness of every ranking including this one. A 200-day moving average used as a trend filter and a 3-day moving average used as a trigger share a name and nothing else. They belong at opposite ends of this table, and right now they sit in the same row.

I said in the last piece that I could not answer this, because the lookback column was empty across all 304,426 SimpleMovingAverage rows. The column was empty, and the reason was a bug of mine. The corpus exporter looked for a top-level lookback field. The window is stored at window.length, so the extractor found nothing and wrote a null for all 4,424,078 indicator rows. Nothing failed loudly, because the export succeeded and the column existed. It was simply always empty, and the value was sitting in the same table the whole time, one column over in the raw parameters.

Read correctly, and with cloned experiments collapsed, it answers the question, and the answer is yes.

This took two corrections to get right, and the first published version of each was wrong. Bucketing every backtest by every window it contained counted crossover strategies twice and produced a 0.261 spread. Fixing that but leaving cloned experiments in place flattened the whole thing to 0.038, with every difference straddling zero, and I published that as a null result. Collapsing clones as well, a 20-day average beats a 200-day one by 0.244 with an interval of [+0.175, +0.310], and all three pairs separate.

So the setting matters more than the type. SimpleMovingAverage sits tenth at 0.613, just below PriceRateOfChange at 0.625, and its neighbours sit within about 0.06 of it. The windows inside SimpleMovingAverage span 0.249, four times that gap, and unlike the neighbours they separate cleanly. Ranking types puts the answer at the wrong level, and the shorter the average the better it did.

3. Train on the outcomes

A strategy paired with what it did is a labelled example. 222,010 of them is a training set, and a rare one: the input is structured rather than prose, the label is a number rather than a judgement, and nobody wrote any of it to be training data.

That is the shape you need to fine-tune a model to write strategies that land in the top decile instead of the middle of the distribution. Every language model can already describe what a good strategy looks like. This would be one that has seen 222,010 attempts and what each of them returned. The obvious hazard is that a model trained to maximise backtest Sharpe learns to overfit backtests, so the label has to be out-of-sample performance rather than the number the corpus already holds.

Corrections

What twice the data revised

The corpus doubled, so I checked what it broke. Three conclusions from my 2025 analysis of 114,549 backtests did not survive the larger sample.

A second result

The sell rule

While the ranking was running I split every backtest containing a buy action on whether it also contained a sell action. Strategies that never sold posted a median return of 36.41% against 19.50% for strategies that did. Nearly double, and it holds in both eras rather than being a 2020s bull-market artifact.

Then I restricted the same split to runs carrying a benchmark. Of the strategies that never sold, only 43.3% beat simply holding the asset, with a median alpha of -0.85. Of the strategies that did sell, 57.0% beat it, at +8.77.

Methodology

How this was measured

The source is a snapshot of the NexusTrade backtest archive taken on 6 September 2026, covering 400,203 scanned documents. Every figure above comes from that snapshot under one filter: the backtest was traded, it ran between 365 and 2,000 days, and it was authored by a person rather than the optimizer.

Each backtest counts once per indicator type, so a strategy using the same indicator in four conditions contributes one observation rather than four. Rankings are medians, not means, because a handful of leveraged outliers move a mean and tell you nothing about the typical strategy. Every median carries a 95% bootstrap interval from 4,000 resamples, and where two intervals overlap I do not claim an ordering between them. Counting is by distinct EXPERIMENT rather than by backtest: cloned and forked portfolios mean the same strategy on the same tickers over the same window appears many times, and counting each copy treats one experiment as a thousand independent observations. Keying on the strategy logic, the tickers and the window keeps 47.8% of eligible runs. Window comparisons additionally count only strategies using a single window of that type, because a crossover strategy holds two and would otherwise be counted in both.

Only aggregate results for built-in indicator types appear anywhere in this analysis. Creator-owned custom indicators, their names, their parameters, their identifiers and their underlying strategy logic stay private, and the results-only table they live in cannot be joined back to any strategy, portfolio or user.

Ask the corpus yourself

Every table in this piece is a question I typed in English. The archive is queryable by the platform's AI agent, which writes read-only SQL against the backtest tables and hands back the rows, so "median Sharpe by moving-average window, minimum 500 backtests" is a sentence rather than a query. Ask which setting of an indicator you already use has done best, whether your rule reads better driving a buy or a sell, or how a type performs once you hold the timeframe fixed.

Ask carefully, though, because the agent answers the question you asked. Put that moving-average question to it plainly and it will bucket every backtest by every window it contains, count each 50/200 crossover twice, and handed me a spread I nearly published. Adding "only strategies that use a single window" changes the answer and is the difference between a finding and an artefact. The corpus removes the work of gathering the data. It does not remove the work of asking a clean question.

Every number here comes from the NexusTrade backtest archive, which is why each row carries its sample size instead of a screenshot of a chart. If you want the anonymized aggregate results to run the ranking yourself, send me the word CORPUS.