How It Works
- Keep a running total of volume: add the whole bar's volume when it closes up, subtract it when it closes down. That total is On-Balance Volume.
- Buy when OBV crosses above its own 20-bar average — buying volume is accumulating faster than it has been recently, which Granville argued tends to lead price.
- Sell when OBV falls back below that average. The comparison is against its own average rather than a fixed level because OBV's raw value depends entirely on where the count began.
Worked example. OBV has been drifting near its 20-bar average when three heavy up days add 2,400 units and lift it clearly above — the strategy buys, even though price has barely moved. Price follows over the next fortnight, and the position is closed when a run of heavy down bars drags OBV back beneath its average.
The Math Behind The Indicators
Everything runs on closing prices of the traded timeframe: P is a close, Pt today's close, and N counts bars — one bar is one candle of that timeframe, so 20 bars on a 1h chart is 20 hours.
- On-Balance Volume (OBV)
- A running tally of volume signed by direction: the bar's entire volume is added when it closes up and subtracted when it closes down. The premise is that turnover reveals conviction — a rally on heavy volume is being accumulated, the same rally on thin volume is not. Only the sign of the price change matters, never its size, so one huge quiet bar counts for less than a small busy one. The absolute level is meaningless (it depends on where the count started); only its direction carries information.
- OBVt = OBVt−1 + {VtPt > Pt−1−VtPt < Pt−10otherwise
- Example: Starting from 0, an up bar on 500 units of volume takes OBV to +500. The next bar closes down on 200 units, bringing it to +300. A third closes up on 900, lifting it to +1200 — buyers are accumulating faster than sellers are distributing.
Example Chart
Example Chart
The Metrics
| Metric | Calculation | What it shows |
|---|---|---|
| Price Change % | change = Plast − P0P0 × 100 The traded market's own close against its first close over the same window, as a percentage. | What the market did while the rule was running — the benchmark every other row here is read against. A rule that made 40% in a market that made 120% lost to doing nothing. |
| Trades | N = count(closed positions) How many positions the rule opened and closed over the window. | The sample behind every other figure, and what the fees are charged on. Two rules with the same return are not the same rule if one took nine trades and the other took nine hundred. |
| Win Rate % | W%n = winsnn × 100 Of the first n trades, how many closed above the cash they opened with after fees. Plotted trade by trade, so the line is the rate so far rather than a final figure. | How often the rule is right, which is not how much it makes. A rule can win a third of its trades and still lead, if the third it wins pays for the two it loses. |
| Cumulative P&L % | PnL%n = n∑i=1(fi − 1) × 100 Each trade's percentage result added up, net of fees. A sum rather than a compounding, so a 10% gain and a 10% loss cancel. | What the rule returned per trade, with position size taken out of it. It answers whether the edge is in the trades themselves, where the equity curve answers what the account did with them. |
| Equity | En = E0 n∏i=1fi The account compounded through every trade — the whole balance goes into the next position. Drawn net of fees as a solid line and gross of them as a dotted one. | The account itself, which is the only figure a reader actually ends up with. The gap between the two lines is what the fees took, and it widens with every trade rather than staying a fixed share. |
| Cumulative Fees | Fn = n∑i=1(Ci φ + Xi φ) Fee charged on the way into each position and again on the way out, at rate phi, on the capital actually committed — so the bill grows with the account as well as with the trade count. | The cost of trading, in the account's own units. It is the one line here that only ever rises, and the one a rule cannot trade its way out of. |
| Rolling Sharpe | Sharpet = mean(rdaily)sd(rdaily) × √365 Mean daily return over its deviation, annualized on a 365-day year because crypto has no weekend. Taken on the account marked to market every bar — open positions included, not just closed ones — and read off at each trade's exit. | Return per unit of the swing it took to get it. It is the heaviest weight in the composite score, because an account that doubled calmly and one that doubled violently are not the same result. |