Conclusion: EWMAC trend following on retail CFDs is not viable after swap costs. This repository documents the complete research process — from successful gross backtesting to the definitive cost analysis that proved the strategy inoperable on CFDs.
Systematic portfolio trading system based on Robert Carver's Advanced Futures Trading Strategies, applied to 10 CFD instruments on Darwinex Zero. Daily timeframe, continuous position sizing via volatility targeting, zero optimized parameters — all values taken directly from published literature.
EWMAC trend following across 10 instruments produces a respectable gross equity curve over 26 years. Sharpe 0.31, CAGR 3.2%, Max DD -28.5%. The system works before costs.
Adding realistic CFD transaction costs (spread + commission + daily swap) reveals the core issue. The orange line shows cumulative swap costs exceeding $100K over 26 years — more than the initial capital. Swap alone turns +$130K gross profit into -$80K net loss.
A "carry-gate" mechanism was developed to penalize positions during high-swap periods. Even at the most aggressive threshold (θ=3%, blocking up to 45% of trading days on some instruments), the best achievable Profit Factor is 1.014 — well below the 1.10 minimum for live trading.
| Scenario | Sharpe | CAGR | Max DD | Profit Factor | Verdict |
|---|---|---|---|---|---|
| Gross (no costs) | 0.314 | 3.2% | -28.5% | 1.07 | System has edge before costs |
| Net (spread + commission only) | 0.216 | 2.1% | -32.8% | 1.04 | Marginal without swap |
| Net (full costs, no gate) | -0.456 | -6.4% | -88.0% | 0.88 | Destroyed by swap |
| Net (carry-gate θ=3%) | 0.118 | 0.6% | -36.0% | 1.01 | Best case: still inoperable |
| Instrument | Net PnL | Note |
|---|---|---|
| USDJPY | +$20,160 | Best performer (positive swap) |
| NIKKEI225 | +$15,133 | Genuine trend follower |
| NASDAQ100 | +$9,691 | Recovers with aggressive gate |
| SP500 | +$3,826 | Marginal positive |
| AUDUSD | +$929 | Breakeven |
| EURUSD | +$56 | Breakeven |
| GOLD | -$4,717 | Swap too high |
| DAX40 | -$5,977 | Trap: looks good gross, fails net |
| SILVER | -$10,496 | Chronic loser |
| GBPUSD | -$10,331 | Double-negative swap |
Carver's framework was designed for futures, where:
- Holding costs are embedded in the price (contango/backwardation), not charged daily
- Transaction costs are a fraction of the daily range
- The edge-to-cost ratio is fundamentally different
On retail CFDs, the daily swap charge creates a structural drag that trend following cannot overcome. The strategy needs positions held for weeks/months to capture fat tails, but every day in a position costs money. This is not a parameter problem — it's a structural incompatibility.
We developed a continuous carry-gate penalty:
This scales down the forecast when swap costs are high relative to volatility. It recovered 0.57 Sharpe units (from -0.46 to +0.12) but the structural gap is too large for any filter to bridge.
Despite the negative conclusion, the codebase implements a complete Carver system:
| Phase | Description | Status | Key Finding |
|---|---|---|---|
| 1 | Data pipeline (10 instruments, 26 years) | ✅ | Yahoo Finance daily OHLCV |
| 2 | EWMAC single-speed (64/256) validation | ✅ | Forecast ≈ N(0,10) confirmed |
| 3 | Multi-speed EWMAC (4 speeds + combination) | ✅ | FDM = 1.04 (Carver reference) |
| 4 | Carry signal investigation | ❌ | Discarded — no improvement over EWMAC |
| 5 | Portfolio + IDM (10 instruments) | ✅ | IDM = 2.17, Sharpe 0.31 gross |
| 6 | Transaction costs (time-varying swap) | ✅ | Swap destroys all edge |
| 6b | Universe filtering + carry-gate | ✅ | Best PF = 1.014 — inoperable |
Where:
- Forecast ∈ [-20, +20], scaled so |mean| ≈ 10
- Daily risk target = Capital × Annual vol target / √252
- Volatility estimated via exponential weighted std (span=25)
core/ - EWMAC forecasts, vol targeting, costs model, carry-gate
backtest/ - Pandas-based daily engine with dynamic capital
config/ - Instrument metadata + Darwinex CFD cost parameters
tools/ - Phase runners, data download, analysis scripts
data/ - Daily OHLCV from Yahoo Finance (not tracked)
analysis/ - Generated charts (not tracked)
images/ - README assets
All from Carver's published work — zero optimization:
- EWMAC forecast scalars from Ch.15 lookup tables
- Forecast Distribution Multiplier (FDM) from Ch.15
- Instrument Diversification Multiplier (IDM) calculated from correlations
- Equal weights (1/N) across instruments and forecast speeds
- 10% position buffer to avoid unnecessary adjustments
- 12% annual volatility target
- Mark-to-market capital — positions scale with equity
SP500 · NASDAQ100 · DAX40 · NIKKEI225 · GOLD · SILVER · EURUSD · USDJPY · AUDUSD · GBPUSD
All via Darwinex Zero CFDs, daily timeframe, 2000–2026.
pip install -r requirements.txt
# Download data (10 instruments)
python tools/download_data.py
# Run gross portfolio backtest
python tools/run_phase5_portfolio.py --save-only
# Run cost analysis with carry-gate theta sweep
python tools/run_phase6b_theta.py --save-only- Always model costs first. A Sharpe 0.31 gross system can be Sharpe -0.46 net.
- Swap is the silent killer on CFDs. It's not the spread or commission — it's the daily financing charge on held positions.
- Trend following needs cheap holding costs. The strategy is designed for futures where cost-of-carry is priced into the contract, not charged as a daily fee.
- Carry-gate is a valid concept but cannot overcome structural cost disadvantages. It's a filter, not a fix.
- Carver's framework is excellent for risk management and position sizing — those components are reusable even if EWMAC on CFDs fails.
- Carver, R. — Advanced Futures Trading Strategies (2023)
- Carver, R. — Systematic Trading (2015)
- pysystemtrade — Reference implementation
MIT