The Cost of Retail Alpha

Retail traders are sold an endless supply of strategies with confident win-rate claims. Most are never tested with the costs a real account pays. The purpose of this study is not to find a profitable system, but to apply one honest standard, realistic costs and an out-of-sample split, to a set of popular and classical ideas, and to report the results plainly, including the ones that fail. Our own two earlier builds are included among the failures.

Our findings on market efficiency and the decay of simple premia are consistent with the academic literature. The contribution here is empirical and practical: a like-for-like teardown of strategies people actually trade, measured on high-quality data, and a clear account of one cost, CFD financing, that quietly decides whether a diversified book is viable for a retail trader.

The dataset is a broker tick feed of bid and ask quotes, covering XAUUSD from 2013 and a universe of 45 instruments (28 foreign-exchange crosses, six metals, and eleven equity indices and energy) from 2018. Total gold history exceeds 400 million ticks. We resample ticks to 5-minute and 1-hour bars, carrying the mean spread per bar, and to daily bars for portfolio tests. The primary evaluation window is 2018 to 2025, where the feed is cleanest.

Every test applies the actual spread from the data. For strategies that hold positions, we add financing. We split results into in-sample (2018 to 2021) and out-of-sample (2022 to 2025), size positions with information available only up to the prior bar, and treat backtesting as an experiment rather than a demonstration. We report the Sharpe ratio, annualized return, and maximum drawdown, and we distinguish clearly between what is measured and what is inferred.

We first tested two strategies of the kind sold to retail traders. The first combines the Supertrend indicator with support and resistance zones. The second is a gold “London session liquidity sweep”, which waits for price to sweep the Asian range and then reverse, and which its source material claims wins 70 to 80 percent of the time at a fixed one-to-two risk-to-reward.

The sweep strategy was tested on gold across the full 2018 to 2025 window. Its real win rate was 33 percent, close to the break-even point for a one-to-two payoff and below it after costs. It lost money in seven of eight years.

Table 1. The promoted 70 to 80 percent win rate does not appear. Costs turn a coin-flip payoff into a steady loss.

Before testing more strategies, we asked whether gold returns contain any linear structure a rules-based system could exploit. They do not. The autocorrelation of returns is close to zero at every horizon we measured, which means neither simple momentum nor simple mean reversion is present. This single fact explains why a trend-following indicator and a reversal method both failed: each was betting on structure that is not in the data.

Table 2. Gold returns behave close to a random walk in linear terms.

We also found a strong intraday pattern at 23:00 UTC, worth roughly 2 basis points an hour with a t-statistic of 6.0. It vanished the moment it was traded properly. Entering at the ask and exiting at the bid reduced it to 0.17 basis points with a t-statistic of 0.53. It was an artifact of the daily re-open, where the spread widens, not a move a trader can capture. This is a useful warning: a large number before costs can mean nothing after them.

Gold and silver move together, with an hourly return correlation of 0.74, which invites a pairs trade. But co-movement is not the same as a stationary spread. The gap between the two has a half-life of roughly 134 days and drifts for months, so fading it simply loses as it keeps widening. A z-score reversion trade returned about minus 16 percent a year with a Sharpe ratio of minus 2.5, negative both in-sample and out-of-sample.

We then tested momentum at the portfolio level across all 45 instruments, since trend-following is documented to work on diversified baskets even when it fails on a single asset. Over this window and universe, it did not.

Table 3. Momentum did not pay here. Only passive long exposure did, and that is beta, not skill. The universe lacks bond futures and 2018 to 2025 was a weak trend regime, so this is not a claim that momentum is dead everywhere.

The only approach that made money was the simplest: hold a diversified basket long, with a filter that cuts exposure when the basket falls below its long-term trend. Before financing, the clean version earned about 6.6 percent a year. This is real, but it is market exposure, not a predictive edge. It made money because gold and equities rose.

It also does not survive the way a retail trader would hold it. A diversified basket has low volatility, so reaching a useful 10 percent volatility requires leverage of about 1.5 times for the clean book and 2.8 times for the full one. On a contract-for-difference account, financing is charged every day on the entire levered position, at roughly the short-term interest rate plus a broker markup, which was above 5 percent a year in 2023 to 2025. That charge is larger than the premium the basket earns.

The clean book keeps a small positive return of about 1.5 percent a year after financing, with drawdowns near 30 percent, which is not worth trading. The full book, which needs more leverage, turns negative. The lesson generalizes: for a retail trader, holding a leveraged diversified position for weeks or months is a fight against the financing rate, and in a high-rate environment the rate wins.

This outcome is specific to leveraged CFDs. On a futures account the economics change in the trader’s favor in two ways. The cost of carry is already embedded in the futures price rather than charged daily on notional, and the cash held as margin earns interest, which offsets much of the drag. This is why diversified trend and risk-premia strategies are run on futures by professionals, not on retail CFDs. An unlevered portfolio of exchange-traded funds avoids the financing charge entirely, but then the strategy is simply low-cost investing, and offers no reason to pay for automation.