Week 01 · 23 August 2026

None of the 4 indicators beat buy & hold once costs were applied.

Tested on 763 S&P 500 and FTSE 350 companies, measured over 1 August 2023 to 21 August 2026 — data the rules had never seen. Buy & hold returned a Sharpe ratio of 1.50 over the same window. MACD, VWAP, EMA produced a negative Sharpe after costs — they lost money net of what it cost to trade them.

Universe

763

S&P 500 & FTSE 350 names

Out-of-sample

3.1 yrs

1 August 2023 – 21 August 2026

Cost assumption

8

bps per side, all in

Benchmark Sharpe

1.50

buy & hold, net

The decay curve

In-sample, out-of-sample, and after costs

Sharpe ratio by indicator: in-sample, out-of-sample, and net of costs — Three readings per indicator. A genuinely useful signal shows three similar bars; the usual result is a tall first bar and a third bar at or below zero.

Results

Every number, in one table

763 companies — MMM, AOS, ABT, ABBV, ACN, ADBE and 757 others — over 17 August 2016 to 21 August 2026, costed at 8 basis points per side.

Indicator performance, Week 01
Indicator	In-sample Sharpe	Out-of-sample Sharpe (gross)	Out-of-sample Sharpe (net)	Trades	Cost paid	Max drawdown	Verdict
RSI	0.12	0.56	0.48	86717	9.4%	-6.4%	WEAKENED
MACD	0.15	-0.35	-0.62	249638	52.3%	-25.0%	FAILED
VWAP	-0.19	-0.30	-0.70	378989	79.4%	-27.1%	FAILED
EMA	-0.15	-0.79	-0.85	60426	12.6%	-30.6%	FAILED
Buy & hold	0.86	1.50	1.50	0	0.1%	-14.3%	benchmark

SURVIVED — positive out-of-sample Sharpe after costs, and beat buy & hold.
WEAKENED — still positive after costs, but lost to buy & hold.
FAILED — out-of-sample Sharpe was zero or negative once costs were applied.

Thresholds are fixed in code, not judged by eye, so the standard cannot drift week to week to suit the result.

Indicator by indicator

What happened, and why

RSI

WEAKENED

Mean-reversion: buy below RSI 30, short above RSI 70, flatten back at 50.

Posted a Sharpe ratio of 0.12 in-sample, 0.56 on data it had never seen, and 0.48 once trading costs were applied — across 86717 trades that paid away 9.4% of notional.

MACD

FAILED

Trend-following: long while the MACD line (EMA12-EMA26) is above its 9-day signal line.

Posted a Sharpe ratio of 0.15 in-sample, -0.35 on data it had never seen, and -0.62 once trading costs were applied — across 249638 trades that paid away 52.3% of notional.

VWAP

FAILED

Trend-following: long while the close is above the 20-day rolling volume-weighted average price.

Posted a Sharpe ratio of -0.19 in-sample, -0.30 on data it had never seen, and -0.70 once trading costs were applied — across 378989 trades that paid away 79.4% of notional.

EMA

FAILED

Trend-following: long while the 20-day EMA is above the 50-day EMA.

Posted a Sharpe ratio of -0.15 in-sample, -0.79 on data it had never seen, and -0.85 once trading costs were applied — across 60426 trades that paid away 12.6% of notional.

Walk-forward check

Did it hold up repeatedly, or once?

A single 70/30 split gives one reading, and one reading can be luck. Each indicator is re-tested across rolling three-year training windows with the following year held out, giving many independent out-of-sample readings instead of one.

The column that matters is not the mean — it is how often the indicator was positive. A strategy averaging 0.4 on the back of one spectacular year and four flat ones is a very different proposition from one that earns 0.4 every year.

Indicator	Windows	Mean Sharpe (net)	Windows positive	Worst window
RSI	6	0.13	50%	-0.65
MACD	6	-0.39	50%	-2.06
VWAP	6	-0.87	17%	-1.96
EMA	6	-0.43	17%	-1.50

Cumulative

Growth of 1.00, net of costs

Equity curves for each indicator against buy and hold — Equal-weighted across the universe, log scale. The benchmark is the grey dashed line.

Sensitivity

How wrong could the cost assumption be?

The obvious objection to any cost-adjusted backtest is that the cost number was chosen by the author. Showing the whole curve is a better answer than defending one point on it. Where each line crosses zero is the highest trading cost that indicator could tolerate and still break even.

Out-of-sample net Sharpe as the cost assumption rises — Out-of-sample Sharpe, net, as the assumed cost rises from zero to 40 basis points per side.

Caveats

What this study cannot tell you

Stated every week, because they do not go away.

Survivorship bias. The universe uses current index membership, so companies delisted or dropped along the way are absent. This flatters absolute returns. It affects the in-sample versus out-of-sample comparison far less, since both halves share the bias — which is why that comparison, not the absolute return, is the headline.
One cost assumption, applied uniformly. Real spreads vary by stock and by day, and widen exactly when you most want to trade. The sensitivity analysis shows how conclusions change across the full range rather than defending a single number.
What is labelled VWAP is a rolling volume-weighted moving average. True VWAP is intraday: it resets each session and cannot be computed from daily bars at all. The daily-chart “VWAP” used across retail trading is the rolling version tested here.
No position sizing, leverage or borrow costs. Every position is the same size, and shorting is assumed free. It is not.

Reproduce this. Every figure above comes from one command against public data: python run_pipeline.py --universe all --week 1. The plain-text report and the underlying data are both published.