Random

What It Does

On each US equity trading day, at 9:30 ET, the strategy generates a random direction signal. With the default settings, that’s a single coin flip - heads opens a long, tails opens a short. The entry is a market order with a bracket attached: a take-profit a fixed number of points above the fill (150 points by default) and a stop-loss a fixed number of points below (50 points by default). If neither bracket fires during the session, the position is force-flattened at 16:00 ET. No overnight risk.

The random seed makes the flip sequence reproducible. The same seed on the same data produces the same trade sequence every run.

Because the entry has no predictive edge, any positive expectancy comes from the asymmetry between take-profit and stop-loss distances. A 150-point target against a 50-point stop skews the reward/risk ratio in the trader’s favor, even though wins happen less often than losses. That shape is the lever available to tune.

Best run against 1-minute or 5-minute data so the 9:30 ET entry and 16:00 ET exit land on real candles instead of a bar close half an hour away.

Ensemble Mode

Setting Ensemble Size above 1 runs N independent random generators in parallel and takes the majority vote as the day’s signal. Each generator draws a random number from 0 to 999; if it’s divisible by 2 it votes long, otherwise short. The strategy trades whichever side wins the vote.

This reproduces the article’s experiments at N = 100, 1000, and 10000. The hypothesis users sometimes bring is that “averaging more random signals” should produce a more reliable edge - the wisdom of crowds applied to coin flips. The empirical answer is no: a thousand coins flipped is still a coin flip. The aggregate behavior is statistically equivalent to a single flip with extra steps.

The optional Skip Band filters out days where the ensemble vote is too close to a 50/50 split. With the band set to [0.485, 0.515] (the article’s value), the strategy only trades when at least 51.6% of generators agree on a direction. Intuition says that filtering down to “high-conviction” signals should improve results. In practice it tends to make things worse: by reducing trade count, it removes shots at the favorable risk/reward asymmetry that was the only edge in the strategy in the first place.

The shipped samples cover all three modes side by side on NQ: a single-generator baseline (N = 1), an unfiltered ensemble (N = 1000, default band), and a filtered ensemble (N = 1000, band [0.485, 0.515]). Comparing them directly: the unfiltered ensemble behaves nearly identically to the single flip (same-ish trade count, same-ish outcome - majority voting among coin flips is still a coin flip). The filtered ensemble takes ~75% fewer trades and flips the result from positive to negative, confirming that filtering to “confident” noise signals removes the favorable-bracket shots that were the only edge.

Key Characteristics

• One trade per US equity trading day, entered at session open
• Direction is a coin flip (or ensemble majority vote) driven by the configured seed
• Point-based take-profit and stop-loss, configurable independently
• Force-flattened at session close if no bracket fires
• No overnight exposure
• Optional N-generator ensemble (up to 10000) with optional skip-band filter

Possible Uses

• Exploring how risk/reward asymmetry shapes outcomes independent of signal quality
• Running as a reference or control alongside a real strategy to separate signal effects from risk-shape effects
• Reproducing the article’s ensemble experiments end to end
• Stress-testing the backtest pipeline
• Comparing tunable parameters (SL/TP combinations, seeds, ensemble sizes, filter bands) under a fixed intraday structure
• Running across instruments to see how the same point-based brackets behave at different price scales

Further Reading

This strategy implements the experiments described in Backtesting Random Trading Signals, including the single-generator baseline, the N-generator ensemble extension, and the probability-band filter variant. The article walks through the motivation, the failed hypotheses, and what the results actually show.