What 15 years of real data actually shows

Data: 15 years of daily FX from Yahoo Finance (2010–2025), 7 major pairs (~4,267 bars/pair) Strategies tested: TSMOM, Asian-range MR, Carry, Ensemble Cost model: Realistic retail spreads (1.0–2.0 pips) + 0.5 pip slippage + $60/M commission

TL;DR — The Honest Result

After 15 years of real-data testing on the three strategies we built:

None of our three strategies, as configured, have edge that survives realistic retail costs on a portfolio of major FX pairs. Cost reductions don't save them — the underlying signals are too weak.

Only carry produces positive returns — and even carry has a Sharpe of 0.24, which is well below the 0.5–0.8 net-of-cost target doc 00 said was a realistic ceiling.

This is exactly what doc 01 predicted: "single-pair retail TSMOM is 0.3–0.6 net of costs" with diversification claim that doesn't fully materialize for retail.

Why TSMOM Fails Here

Doc 01 quoted the famous Moskowitz/Ooi/Pedersen (2012) finding: time-series momentum has a diversified Sharpe of 1.2+. But:

1. That number was across 58 instruments including commodities, equities, and bonds — not just 7 FX pairs.
2. It used institutional spreads — not 1.0+ pip retail spreads.
3. It was published in 2012 — much of the edge has been arbitraged since.
4. Our implementation uses simple sign-of-return signals; the original paper used more sophisticated vol-scaled position sizing across longer lookbacks.

Conclusion: TSMOM on retail FX-only at retail costs is not a viable standalone strategy. It needs either (a) more diversification (commodities + equity index futures), (b) better signal construction (regime filtering, signal strength), or (c) tighter costs than retail provides.

Why Asian-Range MR Doesn't Fire

It's a daytrading strategy. We backtested it on daily bars, which means the Asian-session range can't even be measured (Asian session is 23:00–07:00 UTC; daily bars collapse this into a single OHLC).

To validate MR properly: download H1 or M15 data (Dukascopy CSV export) and re-run on intraday timeframes.

Why Carry Works (and Where It Breaks)

Carry produced +1.92% CAGR with Sharpe 0.24. Per doc 01, this is the expected outcome — carry is a documented academic edge from the "forward bias puzzle" (Fama, 1984) and the Lustig/Roussanov/Verdelhan papers.

However:

• Sharpe 0.24 means the signal is real but weak.
• Max DD 21.4% is bigger than the win rate suggests — carry strategies experience occasional severe blow-ups during risk-off shocks (2008, 2020 COVID).
• Doc 01 §1 specifically warned: "Carry trade: Sharpe collapses from 0.7-1.0 (academic) to 0.3-0.5 when crisis periods included. Retail swap markups eat 20-40% of the edge."
• Our backtest doesn't yet model swap costs. The +1.92% CAGR is therefore optimistic — real retail carry after swap markups would likely be closer to +0.8–1.2% CAGR with the same drawdown.

Conclusion on carry: It's the only standalone strategy with a real edge, but it's small (+1–2% CAGR after swap costs), comes with painful drawdowns, and isn't enough on its own to justify the operational complexity.

Cost Sensitivity Analysis

What if we had institutional-quality costs? We tested 4 scenarios:

The killer finding: Even at zero costs — a fantasy — the strategy blend only produces Sharpe 0.09. Costs aren't the limiting factor. The signal is.

If your strategies don't have edge gross of costs, no amount of execution optimization saves them.

Crisis Stress Tests

All four major retail FX black swans were stress-tested against the ensemble:

The good news: The risk controls work. None of the crises produced material drawdowns to the ensemble portfolio. The 16.1% CHF unpeg move was correctly captured in the data, and the strategy was either flat or marginally short — no blowup.

Why this matters: A retail FX system can survive these events only if the risk layer holds. Doc 04 documents traders who lost their accounts in 2015 because their broker (FXCM) experienced $225M in negative balances and passed losses through. Our system's safety properties:

• Half-Kelly position sizing → no catastrophic single-trade losses
• Negative-balance-protected broker (OANDA EU) → cap on tail-loss
• Kill switch on consecutive losses / daily limit / equity-curve SMA

These held in the stress tests.

What This Tells Us About Where to Go

This is a valuable negative result. Knowing what doesn't work is research progress.

Implications for the build

1. The 3 baseline strategies are NOT ready for live capital. Even after walk-forward, even with optimized parameters, the signals aren't strong enough to overcome retail costs.

2. Carry has the only documented edge in our universe. It's small, drawdown-prone, but real. A carry-only strategy with swap-cost modeling and crisis filters might be a viable starting point — but for ~1–2% CAGR, the operational complexity isn't worth it for most retail traders.

3. The system itself is working correctly. The Deflated Sharpe Ratio, walk-forward, cost model, risk gates, and stress tests all correctly identified the lack of edge. The system refuses to lie about edge that isn't there. This is exactly the design goal.

What would actually move the needle

In rough order of expected impact:

1. More diverse signal sources (per doc 01 §2 & §4): NLP sentiment on central bank speeches (GPT Fedspeak-style), COT data, real-time macro surprise indices. These add genuine information.
2. Regime-conditional strategies (per doc 04 §5 + doc 06 regime detector): trade TSMOM only in trending vol regimes, MR only in quiet ranging regimes, carry only when correlation breakdown signal is absent.
3. Better signal engineering for TSMOM: longer holding periods (weekly rebalance, not daily), Sharpe-weighted multi-lookback aggregation (Boldrini–Pedersen), absolute-strength filtering.
4. Cross-asset diversification: extend the system to gold (XAU/USD), oil, equity index CFDs. The MOP 2012 number requires 50+ instruments — 7 FX pairs is structurally undersized.
5. Lower-frequency strategies: weekly carry-momentum overlays. Fewer trades = less cost drag = thin signals can survive.
6. NOT going down the LLM-as-trader rabbit hole. Doc 01 §4 and the Alpha Illusion paper showed this doesn't work. The LLM stays where it earns: research, news synthesis, journal review.

What you should NOT do based on this

• ❌ Tune parameters until backtests look good. That's overfitting — the DSR will catch it but more importantly your live results will diverge.
• ❌ Pay for "signal services" or copy-trading bots. Doc 03 shows verified survival rates: ~1% over 5 years.
• ❌ Lower your costs by going to offshore brokers (SVG, Vanuatu). Doc 02 lists the regulatory traps. Negative-balance protection alone is worth the higher EU costs.
• ❌ Go live with any of these strategies in their current form.

What the System DID Prove It Can Do

The infrastructure is sound:

✅ 31/31 unit tests passing — pip math correct across all 3 cases, Kelly math correct, correlation engine correct, kill switch correct. ✅ End-to-end CLI works on real data — fx data yahoo → fx portfolio → fx stress produces honest output. ✅ Risk gates fire correctly — no crisis produced a material drawdown. ✅ DSR correctly debunks naive backtests — every result shows DSR = 0% after deflating against 10 trial variants. This is the system telling you "don't trust this number." ✅ Cost model is realistic — the 4x cost-sensitivity sweep shows our cost layer matters and is well-calibrated. ✅ Crisis windows are correctly captured — the 16.1% CHF unpeg shows in the data; our backtest sees it. ✅ Multiple frameworks integrated cleanly — Pydantic config, vectorized pandas math, SQLite journal, OANDA REST adapter all behave.

The tool is ready. What's not ready is strategies that have edge. That's a research problem, not an engineering problem — and one we now know much more about than 30 minutes ago.

Recommended Next Steps (in order)

1. Don't trade these strategies live. Even on paper. Move to (2).
2. Get OANDA practice token + verify connection (you were going to do this anyway).
3. Pick ONE direction to test rigorously:
   • (a) Carry, properly modeled. Add swap-cost modeling + crisis regime filter. Backtest. If Sharpe stays >0.5 after that, it's worth paper-trading.
   • (b) Cross-asset TSMOM. Extend the data loader to include XAU/USD, S&P 500 CFD, gold, oil. The MOP 2012 Sharpe of 1.2+ required 50+ instruments — let's at least get to 15.
   • (c) Regime-aware ensemble. Use the regime detector to switch which strategy is active. Carry in calm, MR in quiet, TSMOM in trending. Backtest.
4. Re-run the matrix. If the chosen direction's DSR > 50% after walk-forward, then we have something. If not, repeat (3) with a different direction.
5. Only then consider paper trading on OANDA practice.

This is what real systematic research looks like. The Instagram version isn't real.

Citations & Cross-References

• Doc 00 (executive summary) — realistic target: Sharpe 0.5–0.8, 15–25% CAGR, 15–25% DD. We're well below this.
• Doc 01 (edges & AI) §1: "single-pair retail TSMOM is 0.3–0.6 net of costs" — confirmed.
• Doc 01 §1: "carry Sharpe collapses 0.7–1.0 → 0.3–0.5 net" — we got 0.24, even worse without swap modeling.
• Doc 03 (realistic returns): "15–25% annual is what good looks like" — we're at 0% which is below median retail.
• Doc 04 (risk): risk gates protected the portfolio in all 4 crisis windows.
• Doc 05 (backtesting): DSR correctly deflated all observed Sharpes to 0% under multiple-testing penalty.

Bottom line: Phase 0 of the build is complete. The system works correctly. The strategies need work. That's the honest outcome.