A rigorous, honest backtest of the Volume Profile / Market Profile method (Steidlmayer / Dalton): it codes the objective rules — POC reversion, Edge-to-Edge, Value-Area breakout, volume exhaustion and the 80% Rule — measures per-trade expectancy after costs on up to 33 years of data, validates everything out-of-sample (walk-forward), and then tries hard to falsify the result.
The question is not "can I draw nice levels?" but "is there a real, standalone economic edge once you account for costs, overfitting and plain long exposure?" The answer, told without spin, is below.
- A real but modest edge exists on the long side of liquid indices. Out-of-sample, every strategy is profitable on QQQ and SPY (QQQ Edge-to-Edge: Profit Factor 2.06; QQQ Volume exhaustion: expectancy +1.80%/trade).
- "Volume reading" beats the profile geometry. Buying exhaustion / no-supply (new lows on below-average volume) was the most robust tactic OOS on index ETFs (PF 1.5–2.2). Requiring a volume signal candle lifts QQQ Edge-to-Edge from PF 1.14 → 1.89.
- 🇧🇷 It fails on trending Brazilian single names. PETR4 and VALE3 produced negative expectancy on almost everything — exactly what mean-reversion theory predicts in trend.
- Under falsification it does not survive as a standalone strategy. Volume permutation, price-only ablation, random-entry controls, excess-return and bootstrap tests show the gain is largely long exposure to assets that rose. No sleeve generates alpha over its own exposure, beats the risk-free rate at 1% risk, or has a Profit Factor whose 95% CI excludes 1.0. Only SPY shows a statistically real volume signal — small and fragile in-sample.
- Two famous "legends" do not hold up. The 80% Rule does not traverse 80% of the time and loses after costs; day-types do not predict continuation (if anything, they revert).
One-sentence verdict: Volume Profile is a legitimate context lens and the volume filter adds real selectivity, but there is no robust, standalone economic edge — it is not a "secret formula".
Full 19-page technical report: output/volume_profile_study.pdf
- Background
- Data & methodology
- Results
- Conclusions
- Reproduce it
- Project structure
- Reference document
- Disclaimer
A Volume Profile is a volume-by-price histogram: it shows where, over a chosen window, trading actually happened. Three levels summarize it — the POC (Point of Control, the most-traded price), and the Value Area high/low (VAH/VAL, the band holding ~70% of the volume). High-volume nodes (HVN) act like magnets where price "sticks"; low-volume nodes (LVN) are gaps it "slips" through. The method's promise is that these levels frame mean-reversion (fade the edges back to the POC) and breakouts (acceptance outside the Value Area).
The composite profile of the last 120 SPY sessions. Price oscillates inside the Value Area and stalls near high-volume nodes — the geometry the strategy trades around.
The original method is intraday (30-minute blocks, Initial Balance, acceptance in two TPOs). Free intraday history is short (Yahoo gives ~60 days of 30-minute bars), so a faithful 20-year intraday backtest is impossible with free data. This study therefore uses a hybrid scope: a long daily backtest with a rolling composite profile (where the statistics live), plus the recent intraday sample used only to sanity-check the 80% Rule.
| Source | Yahoo Finance via yfinance, downloaded and cached automatically as Parquet |
| Instruments | SPY, QQQ (US) · PETR4, VALE3, BOVA11 (Brazil / B3) |
| History | Daily OHLCV, per-instrument from inception (SPY 1993, QQQ 1999, PETR4/VALE3 2000, BOVA11 2009) through June 2026 — up to 33 years |
| Strategies | POC reversion (REV), Edge-to-Edge (E2E), Value-Area breakout (BRK), Volume exhaustion / no-supply (EXH), and the intraday 80% Rule |
| King metric | Per-trade expectancy after costs; also Profit Factor, Sharpe, Sortino, max drawdown, payoff, win rate |
| Costs | Per-market commission + exchange fee + slippage (US and BR cost models), applied round-trip on every trade |
| Validation | Walk-forward (8-year train → 3-year test): parameters are optimized only on past data and measured on the unseen future, so reported numbers are genuinely out-of-sample |
| Falsification | Volume permutation (500×), price-only ablation, random-entry control (500×), excess return vs exposure & risk-free, and a 10,000-sample bootstrap CI on the Profit Factor |
All figures and tables below come straight from the pipeline (scripts/run_experiments.py →
scripts/build_report.py); nothing is hand-edited.
Every strategy was walk-forward validated on every instrument, long-only, with costs. The pattern is sharp: the edge concentrates on liquid US indices and disappears (or inverts) on trending Brazilian single names.
| Strategy | SPY | QQQ | PETR4 | VALE3 | BOVA11 |
|---|---|---|---|---|---|
| POC reversion | 1.13 | 1.63 | 0.90 | 0.99 | 0.85 |
| Edge-to-Edge | 1.41 | 2.06 | 0.95 | 0.34 | 0.66 |
| VA breakout | 1.38 | 1.58 | 0.99 | 0.96 | 1.12 |
| Volume exhaustion | 1.51 | 1.94 | 0.82 | 1.25 | 2.16 |
Out-of-sample Profit Factor. Green > 1 is profitable. The US column is consistently green; the PETR4/VALE3 columns are not.
The whole point of walk-forward is to catch overfitting: parameters chosen on the training window are tested on the next, unseen window. The flagship sleeves keep a positive expectancy across most folds rather than collapsing the moment they leave the optimizer's data.
Per-fold expectancy in R, training (grey) vs out-of-sample (blue). The edge degrades — as it always does — but does not vanish, which is the bar a tradable signal must clear.
Combining only the sleeves that stayed positive out-of-sample (SPY Edge-to-Edge, QQQ VA breakout, BOVA11 Volume exhaustion) gives a low-exposure stream with a strikingly shallow drawdown.
| Portfolio (OOS-validated sleeves) | CAGR | Sharpe | Max drawdown | Exposure |
|---|---|---|---|---|
| 1× | 1.42% | 0.54 | −6.7% | 52% |
| 3× | 4.11% | 0.54 | −19.3% | — |
| 5× | 6.57% | 0.54 | −31.0% | — |
Because the system is in the market only ~half the time, leverage trades return for drawdown at a roughly constant Sharpe. There is no free lunch in the risk dial.
Honesty check: a low-exposure, mean-reversion system will not out-return simply owning the index across a 30-year bull market. What it offers is a far smoother ride.
Buy & hold SPY compounded at 8.8%/yr but suffered a −56% drawdown; the VP portfolio returned less but drew down a single-digit percentage. Different objective, different product.
The flagship SPY Volume-exhaustion edge is real but thin — it survives realistic costs and dies under high ones.
| Round-trip cost | Profit Factor | Expectancy/trade |
|---|---|---|
| 0.01% | 1.78 | +0.83% |
| 0.11% | 1.66 | +0.73% |
| 0.21% | 1.55 | +0.63% |
| 0.41% | 1.35 | +0.43% |
| 0.81% | 1.02 | +0.03% |
Above ~0.8% round-trip the edge is gone. This is why the method only makes sense on cheap, liquid instruments.
The most interesting positive result: requiring a volume spike (a "signal candle") before taking an Edge-to-Edge entry sharply improves QQQ — until the filter becomes so strict it starves the sample.
| Volume requirement (× avg) | QQQ trades | QQQ Profit Factor |
|---|---|---|
| 0.0 (no confirmation) | 75 | 1.14 |
| 1.2× | 54 | 1.63 |
| 1.5× | 29 | 1.89 |
| 2.0× | 10 | 0.60 |
Volume confirmation adds genuine selectivity, but over-filtering (≥ 2×) leaves too few trades to trust. The signal is in the volume, not the geometry.
Two of the method's most-repeated claims do not survive contact with the data.
| Day-types do not predict continuation | The 80% Rule does not traverse 80% |
|---|---|
 |
 |
Left: forward returns by day-type — "bearish" days bounce more than "bullish" ones, the opposite of the promise. Right: in the recent 30-minute sample the Value Area is fully traversed only 27–67% of the time, never the advertised 80% (small samples, indicative only).
This is the section most backtests skip. With the configuration frozen (not re-optimized) and a fixed seed, every sleeve faced six adversarial tests.
| Volume permutation (Test 1) | Bootstrap CI on Profit Factor (Test 5) |
|---|---|
 |
 |
The verdict table is humbling and honest:
| Sleeve | Real PF | Shuffle p | Beats price-only? | Random-entry p | Alpha vs exposure? | Beats risk-free? | 95% CI excludes 1? | Passes all? |
|---|---|---|---|---|---|---|---|---|
| SPY | 1.66 | 0.002 | ✅ (+0.40) | 0.028 | ❌ | ❌ | ❌ [0.99, 2.83] | No |
| QQQ | 1.24 | 0.080 | ✅ (+0.22) | 0.300 | ❌ | ❌ | ❌ [0.71, 2.21] | No |
| PETR4 | 0.78 | 0.840 | ❌ | 0.908 | ❌ | ❌ | ❌ [0.50, 1.19] | No |
| VALE3 | 0.96 | 0.902 | ❌ | 0.704 | ❌ | ❌ | ❌ [0.65, 1.41] | No |
| BOVA11 | 1.06 | 0.112 | ✅ (+0.19) | 0.416 | ❌ | ❌ | ❌ [0.62, 1.79] | No |
Only SPY's volume signal is statistically real (shuffle p = 0.002, ablation positive, random-entry
p = 0.028) — yet it still fails the economic tests: no alpha over plain exposure, it does not beat
the risk-free rate at 1% risk, and its Profit-Factor confidence interval includes 1.0. The "edge" is,
to a large extent, long exposure to assets that went up.
- There is a small, real signal in volume on liquid US indices — strongest in the exhaustion / no-supply read and in the volume signal-candle filter — but it is modest and concentrated.
- The geometry alone (POC, Value Area, day-types, 80% Rule) is mostly folklore. The value comes from volume reading and from using the profile as context, not from the levels as predictions.
- As a standalone system it has no robust economic edge. It does not beat buy & hold on return, does not generate alpha over its own exposure, and its Profit Factor is not statistically distinct from 1.0 after bootstrapping.
- Where it shines is risk: a diversified, low-exposure portfolio delivered single-digit drawdowns versus the index's −56%. That is a risk-management property, not a money-printing one.
The professional takeaway: treat Volume Profile as a lens, demand volume confirmation, trade only cheap and liquid instruments, and keep expectations modest.
The first run downloads and caches the data; later runs are offline. Generated figures and the PDF report are committed, so you can read the study without running anything.
# Windows / PowerShell (Python 3.12+)
py -3.12 -m venv .venv
& ".venv\Scripts\Activate.ps1" # if blocked: Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
pip install -r requirements.txt
pip install -e .
pytest -q # 9 tests
python scripts/run_experiments.py # downloads + caches data, writes output/results.pkl
python scripts/run_falsification.py # optional: writes output/falsification.pkl
python scripts/build_report.py # renders output/figures/*.png and the PDF report# macOS / Linux
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt && pip install -e .
pytest -q
python scripts/run_experiments.py && python scripts/run_falsification.py && python scripts/build_report.pysrc/vptrading/
data/ download + Parquet cache of OHLCV (yfinance); daily history + recent 30-minute bars
core/ Volume Profile math (POC, Value Area 70%, HVN/LVN), rolling composite, day-types
strategies/ POC reversion, Edge-to-Edge, VA breakout, Volume exhaustion, and the 80% Rule
backtest/ cost models (US/BR), metrics (expectancy, Sharpe, Sortino, drawdown, profit factor),
and the event-driven simulation engine
optimization/ grid search + walk-forward (in-sample -> out-of-sample)
analysis/ claim studies (day-types, volume divergence, absorption) and the falsification suite
reporting/ figure generation (matplotlib) and the PDF report (reportlab)
scripts/ run_experiments.py · run_falsification.py · build_report.py
tests/ unit tests for the profile math, metrics, and engine
output/figures/ the 19 figures used in the report and this README (committed)
volume-profile-strategy.md — the strategy itself, the math (Value Area
70%, POC, HVN/LVN), the day-types, and a critical evaluation separating what is methodologically sound
from what is trading folklore.
This is an educational and research project. Nothing here is investment advice. Retail day trading has a negative base rate (≈97% of participants lose money — Chague & De-Losso, USP/FGV), and leveraged trading carries the risk of total loss. Past performance does not guarantee future results.
Licensed under the MIT License.