TL;DR: the answer box
- Bet with the drift and most coins made money, 2007 to 2026. 97.7% of random ES longs made money, 93.9% after costs. Random ES shorts: 2.5%.
- Almost half cleared a profit factor of 1.3. 48.9% of random ES longs had a profit factor of 1.3 or more.
- The typical coin still lagged buy and hold. The median random ES long made $180,962, 66% of buy and hold's $273,725, at a Ret/DD of 2.98 against 4.53.
- Natural gas's fall was mostly the roll. The price fell $29,430 on one contract. The futures chart fell $168,920. 82.6% was the roll.
- Use the coin as a ruler. An ES RSI(2) rule averaged $768 a trade. Random longs with the same hold averaged $266. 34.6% of the rule was drift.
How I tested random entries
Here is the question. Take away every indicator and let a coin decide when to trade. Does the market still pay you? And if it does, what is it paying you for?
I spent years hunting for the best indicator, and lost roughly $270,000 early to overfitting and leverage. A coin is the cheapest way to find out how much of a backtest the market hands out for free. Think of it as a ruler. Anything your strategy earns has to be measured against it.
The coin. Each random strategy has two settings:
- Entry chance: how often the coin says buy. 10%, 25%, 50%, 75% or 90% of days.
- Hold: how long each trade stays open. 1, 3, 5, 10 or 20 bars (one bar is one trading day).
The coin is read at each close. If it says buy, the trade buys the next open. If a trade is already open, that day is skipped. One position at a time.
Long only and short only, kept apart. Each setting runs 200 seeds on each side. That is 5,000 strategies per market per side, 40,000 in all. Every seed is recorded, so every strategy can be rebuilt exactly.
The markets. ES (E-mini S&P 500 futures), natural gas futures and Japanese yen futures, one contract each, plus SPY with $35,000 in whole shares. Daily bars from 2007-01-03 to 2026-07-01, except SPY, which ends on 2026-06-12.
The fills: Execution rule v2. The trade buys the next open after the coin says buy. It sells at the open 5 bars (or whatever the hold is) after it bought. No costs by default, with a costed view beside it: $4.00 a round turn plus one tick of slippage a side on futures. The yen has no costed view, because its engine row has no tick size.
That strategy is real, and it is one of the 40,000. Over the full window it took 567 trades and won 63.1% of them. It made $161,388 on one contract, with a worst drop of $58,238. No indicator. No idea. Just a long bias and a coin.
Can a coin flip trading strategy make money?
On these four markets, yes, if it bets in the right direction. On the side of the drift, 91.2% to 97.7% of random strategies made money. On the other side, 2.5% to 8.3%.
Drift is the direction a market tends to move over many years. ES and SPY drifted up. Natural gas and the yen drifted down.
The picture gets clearer when you draw every strategy. Below are 200 random longs and 200 random shorts on ES, all with the same coin settings.
Every one of the 200 longs ended in profit. The worst made $27,938. Not one of the 200 shorts did. Their median ended at -$193,038.
Now look at win rates. Random ES longs won 58.9% of trades at the median, with a median Ret/DD of 2.98 (Ret/DD is profit divided by the worst dollar drop). Random ES shorts won 40.7%. No signal did that. The market did. In all eight market and side groups, the side with the higher win rate was the side that made money. So a high win rate on a drifting market does not, on its own, prove skill.
That is the idea behind a random-entry baseline. LuxAlgo describes it as swapping a system's entry for a coin flip and keeping everything else the same. This study adds the split that matters: long against short, market by market.
Do random strategies cluster around a profit factor of 1.1 to 1.3?
Partly, and that is the problem. Profit factor is the money won divided by the money lost. Above 1.0 makes money. A common claim is that random strategies on ES land between 1.1 and 1.3.
The median random ES long had a profit factor of 1.29. 39.2% of them landed between 1.1 and 1.3. The middle half ran from 1.17 to 1.47, and the top 5% reached 1.72 or more.
So how many random traders would have looked good?
48.9% reached a profit factor of 1.3 or more. Only 2.3% lost money. A profit factor of 1.3 is the line this study drew for a backtest that looks good. On this evidence, a coin got there about half the time.
Is natural gas's short bias real, or is it the roll?
It is real money for a futures trader. But most of it is not the price going down.
A futures contract expires. To keep a position, you sell the old contract and buy the next one. That swap is the roll. When the next contract costs more than the old one (traders call this contango), a long holder pays up at every roll. A continuous futures chart is "back-adjusted": it glues the contracts together and books that cost as a fall in price.
The EIA's Henry Hub spot series shows the physical gas price at $6.55 per million Btu in January 2007 and $3.15 in June 2026. Down, but nowhere near what the futures chart shows. So I split the move into the price and the roll.
The front-month price fell from 6.163 to 3.220 dollars per million Btu. That is $29,430 on one contract. The back-adjusted contract fell $168,920. The roll was $139,490 of it.
Does that make the short bias fake? No. A short who held through the rolls really collected that money, because the short side is paid the roll the long side pays. But "natural gas wants to go down" is the wrong lesson. The right one is "natural gas futures charged longs to stay in". One bet depends on the price. The other depends on the shape of the futures curve.
ES runs the other way. Its roll cost a long holder $31,325, so ES's drift up is the price itself.
One note on method. The roll split uses a second data export, so it is read as a share, not added to the study's dollar results.
Was the yen a neutral market?
No. The yen was meant to be the control, the market with no drift. Random yen shorts made money 91.2% of the time. Random longs, 8.3%. 57.1% of the yen's back-adjusted fall was roll. The rest was the yen weakening, which the Federal Reserve's yen series shows on the spot rate.
The Australian dollar was the planned neutral market, but the data here has no Australian dollar futures. I cannot tell you it was neutral, because I could not test it. I can tell you the market standing in for it was not.
Does market bias change over time?
Its strength did. Its direction held on all four markets.
The yen's short bias was weak in the first decade: 63.5% of random shorts were profitable. In the second decade, 96.7%. Natural gas went the other way. Its shorts fell from 95.7% to 85.6%, and random longs rose from 4.0% to 14.5%.
The random test shows which way each contract drifted over the window tested. It does not promise the next decade.
Does more exposure mean more profit?
On ES, yes, and the risk-adjusted score rose too. Exposure is the share of days you hold a position. More time in a market that drifts up should pay more. What matters is whether it paid more per dollar of risk.
From a 10% to a 90% entry chance, the median random ES long went from $93,050 to $218,056. Its median Ret/DD rose from 1.94 to 3.28. Time in the market went from 33.2% to 81.8%.
One caution. Over a fixed window, Ret/DD rises with trade count, because profit adds up trade by trade and the worst drop does not. That is why I compare cells with each other, not a cell against a rule that trades less often.
Hold length mattered as much as the coin, and costs hit the short holds hardest.
One-bar ES longs were 90.6% profitable before costs and 74.5% after. With a 20-bar hold, 99.9% survived costs. Short holds churn, and churn pays the broker.
Can a coin flip beat the market?
The median coin flip did not. It rode the same drift as buy and hold, just with fewer days in the market.
On ES, buy and hold made $273,725 on one contract, with a Ret/DD of 4.53. The median random long made $180,962, with a Ret/DD of 2.98. It earned 66% of the profit while in the market 69.2% of the time. Only 7.1% of random ES longs made more than buy and hold, and 13.8% beat its Ret/DD.
Buy and hold had its own cost. The worst drop on one ES contract was $60,375, more than a whole $35,000 account: a fall of 114.35% from the account's peak. One contract on $35,000 is leverage.
How much of a real strategy is drift?
About a third, in the two rules I checked. This is the part you can use.
I set two fixed rules beside the coins. RSI(2) is a 2-day Relative Strength Index, a 0 to 100 gauge of how hard price moved in the last two days. The ES rule buys the next open after RSI(2) closes below 10. The natural gas rule sells short after it closes above 90. Both hold 5 bars, like one of the random cells.
The ES rule averaged $768 a trade over 245 trades, with a 62.0% win rate and a Ret/DD of 4.04. Random longs with the same 5-bar hold averaged $266. So 34.6% of the rule's average trade came just from being long ES for five bars. The rule's edge over random was $502 a trade.
Is that edge real, or just a lucky draw? Line the rule up against all 1,000 random longs with the same hold.
0.0% of the 1,000 random longs matched the rule's $768. The best one reached $753. The natural gas rule tells the same story: 35.9% drift, $371 a trade of edge, and 0.1% of 1,000 random shorts matched its $579.
This is why "trade with the market's bias" is half a lesson. Siding with the bias gets you the drift. It does not tell you whether your rule adds anything on top. A random-entry test on the same side and the same hold does.
How many random strategies are enough?
For the three groups I checked, a few hundred. Those were ES longs, natural gas shorts and yen shorts, each sampled from its fixed pool of 5,000. At 500 strategies, the middle 90% of samples ran 96.6% to 98.6% for ES longs, 95.2% to 97.8% for natural gas shorts and 89.2% to 93.0% for yen shorts. For each, I drew 1,000 random samples of each size from its 5,000 strategies and measured how much the share profitable moved.
With 100 strategies, the middle 90% of the 1,000 samples put ES longs between 95.0% and 100.0%. The yen shorts, the weakest bias, spread wider: 87.0% to 96.0% at 100 strategies. For these three groups, you did not need 40,000 coins to pin the share down. You need enough to make the range narrow.
The honest limits
- Four markets and one window, 2007 to 2026. A different window gives different shares, as the two decades show.
- No Australian dollar data. The yen replaced it, and the yen trended.
- The headline numbers have no costs. The costed view is in the reference table. The yen has none.
- The roll split comes from a second data export and is read as a share.
- The two RSI(2) rules are fixed comparisons, not tuned and not recommendations.
What this means for you
- Before you trust a backtest, run the coin on the same side. In both rules here, random entries with the same hold earned about a third of the rule's average trade: 34.6% on ES and 35.9% on natural gas.
- Compare average trades, not totals. A total mixes the edge with the trade count. The ES rule took 245 trades and made $188,162. The median random long with the same hold took 717 and made $180,569. On ES, a random long with a 5-bar hold averaged $266. Your edge is what you earn above that.
- On futures, check the roll. On natural gas it was 82.6% of the back-adjusted fall, and it pays a short only while the curve stays in contango.
- Re-run it by decade. The yen went from 63.5% to 96.7% of random shorts profitable between two decades.
- Then test the part that is yours. Comparing a strategy against random, period by period, is what AlgoChef is built to do.
Are markets random at all? That is its own study: why markets are not random. And for what combining several strategies did, see the Ray Dalio diversification study.
Reference tables
Table 1. Long only against short only, every market
| Market | Side | Profitable | After costs | Median net profit | Median Ret/DD | Median win rate |
|---|---|---|---|---|---|---|
| ES | Long | 97.7% | 93.9% | $180,962 | 2.98 | 58.9% |
| ES | Short | 2.5% | 1.1% | -$181,350 | -0.82 | 40.7% |
| SPY | Long | 97.6% | 95.6% | $45,793 | 2.31 | 59.2% |
| SPY | Short | 2.6% | 1.5% | -$45,651 | -0.73 | 40.8% |
| NG | Long | 4.0% | 2.0% | -$135,575 | -0.75 | 46.4% |
| NG | Short | 96.5% | 91.8% | $137,755 | 2.24 | 53.3% |
| JY | Long | 8.3% | no costed view | -$43,244 | -0.53 | 46.1% |
| JY | Short | 91.2% | no costed view | $43,281 | 0.98 | 53.7% |
5,000 strategies per row, 2007 to 2026, Execution rule v2. Futures one contract, SPY $35,000 in shares. Ret/DD is net profit over the worst dollar drop.
Table 2. Two RSI(2) rules against random entries
| Rule | Trades | Win rate | Average trade | Random average trade | Drift share | Ret/DD |
|---|---|---|---|---|---|---|
| ES long, RSI(2) below 10 | 245 | 62.0% | $768 | $266 | 34.6% | 4.04 |
| NG short, RSI(2) above 90 | 260 | 55.0% | $579 | $208 | 35.9% | 5.79 |
RSI(2) read at the close, buy (or short) the next open, sell the open five bars later. Random average trade: mean of 1,000 random strategies, same market, side and hold. One contract, no costs.
The coin did not find an edge. It found the drift, and part of the drift was the contract. Measure both before you call anything yours.
Everything here is research, not investment advice. It is a backtest on historical data with the assumptions stated above. It is not a recommendation to buy or sell anything, and past results do not predict future ones.
I test one strategy everyone repeats, properly, and leave the parts that failed in. StatOasis.com/Overfit
