Executive summary — the honest brief before you risk one euro

This chapter synthesizes the core research into a single honest brief: what the math actually shows about retail FX, where systematic methods help, and the realistic performance bands to anchor your expectations.

TL;DR (read this if nothing else)

1. "Bulletproof" doesn't exist. 68–89% of retail FX accounts lose money in any 12-month window (ESMA-disclosed numbers from real brokers). Average account lifespan: ~4 months. Only ~1% sustain profit over 5+ years.
2. LLMs are not good at price prediction. JP Morgan shut down their deep-learning FX execution model in Oct 2023 — if their quant team can't make it work with prime-broker spreads and tick data, retail "AI signal" services are noise. The 2025 "Alpha Illusion" paper formally debunks the LLM-trading-agent hype cycle.
3. Where AI/ML genuinely helps retail is execution, risk, regime detection, news/sentiment ingestion, and journaling — not generating alpha from price charts.
4. Honest performance target for a serious 12-month build: Sharpe 0.5–0.8 net of costs, 15–25% annual return, 15–25% drawdown. Anything more in backtest = overfitting.
5. Capital reality: sub-€10k = hobby (costs > realistic edge). €25–100k = supplementary. €250k+ = where 15–25% returns produce livable money without ruinous leverage.
6. Recommendation: Yes, build it — but as a research + risk cockpit + paper-trading harness first, with live capital deferred 6+ months until forward-test results justify it. Treat it as a 12–18 month project, not a 6-week sprint.

The Hard Truths (synthesized across all five reports)

1. The math of edge decay

• Carry trade: academic Sharpe 0.7–1.0 collapses to 0.3–0.5 when crisis periods are included, then retail swap markups eat another 20–40% of that.
• Time-series momentum (Moskowitz/Ooi/Pedersen) is the strongest published FX edge, but the famous 1.2+ Sharpe is diversified across 58 instruments. Single-pair retail TSMOM: 0.3–0.6 net of costs.
• Any single-strategy single-pair retail edge that survives realistic costs is small. Real systems stack 3–6 weak edges across multiple pairs/timeframes.

2. Win-rate ≠ profitability

• Top systematic FX/CTA traders run 35–48% win rates with 2:1–3:1 R-multiples. High win-rate strategies (70%+) hide left-tail blowups.
• FundedNext data: 41% of paid CFD traders have win rates under 50%. Edge comes from R, not from being "right."

3. Drawdown is the silent killer

• A 50% drawdown requires +100% to recover. A 30% DD needs +43%.
• Even Sharpe-1.0 systematic traders see 25–45% drawdowns on multi-year samples. If a backtest shows 80% return / 5% DD: it's wrong, not magic.

4. Sharpe ratios lie at small sample sizes

• A backtest Sharpe of 1.5 over 2 years has a 95% CI of roughly 0.7–2.3 (Lo, 2002). Almost meaningless.
• The Deflated Sharpe Ratio (Bailey & López de Prado, 2014) shows that testing 45 variants gives you a "Sharpe > 1" by chance. Most retail backtests are this.
• Minimum track record length at typical retail Sharpe is ~3 years of live data — not tradition, math.

5. Black swans are real and uncompensated

• CHF unpeg 15-Jan-2015: FXCM took $225M in negative client balances; Alpari went bankrupt. EU brokers now have negative-balance protection — non-EU brokers don't.
• GBP flash crash 7-Oct-2016 23:07 GMT: 6% move in 2 minutes during Tokyo open. Stops slipped 100–200 pips.
• Weekend gaps, peg breaks, and broker-specific quote dislocations are the tail-risk you cannot hedge inside a single FX account.

What's Actually Worth Building

Based on cross-referencing all five docs, the system that's defensibly +EV for a retail trader looks like this:

┌─────────────────────────────────────────────────────────────┐
│  LAYER 1 — RESEARCH ENGINE (LLMs shine here)                │
│  • Daily macro brief: CB calendars, COT, DXY, yield spreads │
│  • Central bank speech parser (GPT Fedspeak-style)          │
│  • News sentiment per-pair                                  │
│  • Regime detector (vol clustering, correlation breakdown)  │
└─────────────────────────────────────────────────────────────┘
                          ↓ feeds
┌─────────────────────────────────────────────────────────────┐
│  LAYER 2 — STRATEGY LIBRARY (rules-based, ML-augmented)     │
│  • TSMOM on majors + crosses, multiple lookbacks            │
│  • Mean-reversion on Asian-session ranges                   │
│  • Carry-trade overlay (small allocation)                   │
│  • Event/news avoidance filter                              │
│  → ML role: feature engineering, vol forecasting, sizing    │
│  → NOT: predicting next bar's close                         │
└─────────────────────────────────────────────────────────────┘
                          ↓ sized by
┌─────────────────────────────────────────────────────────────┐
│  LAYER 3 — RISK COCKPIT (the part that saves you)           │
│  • Half-Kelly position sizing with per-pair pip-value calc  │
│  • Correlation budget (n_eff cap of ~3 effective trades)    │
│  • Daily/weekly/monthly loss limits → kill-switch           │
│  • Pre-news/pre-weekend auto-flatten rules                  │
│  • Equity-curve SMA filter (stop trading when underwater)   │
└─────────────────────────────────────────────────────────────┘
                          ↓ executed via
┌─────────────────────────────────────────────────────────────┐
│  LAYER 4 — EXECUTION (broker-native API)                    │
│  • OANDA v20 REST for primary execution                     │
│  • IBKR as multi-asset insurance                            │
│  • Order audit log → journal                                │
└─────────────────────────────────────────────────────────────┘
                          ↓ measured by
┌─────────────────────────────────────────────────────────────┐
│  LAYER 5 — JOURNAL + REVIEW (LLMs shine again)              │
│  • Every trade logged with regime + reasoning               │
│  • Weekly LLM review for behavioral leaks                   │
│  • Deflated Sharpe / PSR computed continuously              │
│  • Walk-forward retrain schedule                            │
└─────────────────────────────────────────────────────────────┘

Recommended stack

Hard "no" list

• MT4/MT5 as the trading core (use only for visual chart cross-checks)
• Any broker registered in SVG, Vanuatu, Belize, Marshall Islands, Seychelles, Mauritius
• Copy-trade / signal-service subscriptions
• "AI trading bot" SaaS — every one tested in the Alpha Illusion paper failed
• Going live with <6 months of forward-test data
• Risking >1% per trade until 100+ live trades logged

Dutch-Specific Watch-Outs

• Tax (from doc 02): Active leveraged FX trading is exactly the profile the Belastingdienst reclassifies from Box 3 (wealth, ~36% deemed yield) to Box 1 (income, up to 49.5%). No bright-line threshold. Keep clean per-trade exports from the broker API from day one.
• AFM oversight comes via MiFID passporting — OANDA Europe (Malta), IBKR Ireland, Pepperstone Cyprus all legitimate.
• Negative balance protection is mandatory for EU-regulated brokers since 2018 ESMA rules. Don't trade with anyone who doesn't have it.

Realistic Targets & Milestones

If you commit to this seriously, here's what good looks like:

Total elapsed before full-size live trading: ~12–18 months. Anyone selling you a shortcut is selling you survivorship bias.

Honest go/no-go questions

Before committing real capital — or weeks of build time — answer these:

1. Capital available for this — and capital you can afford to lose entirely? (Determines whether this is rational at all.)
2. Time horizon — are you willing to spend 12+ months before scaling to real size, or are you hoping for income in 3 months? (If the latter: stop now.)
3. Discretionary or fully systematic? — fully systematic is harder to build but psychologically easier to run. Discretionary trading is the opposite.
4. Pairs/timeframe preference — major pairs on H1–H4 is where the edges are documented. Scalping EUR/USD on 1m is a retail-killer.
5. Existing trading experience — if zero, the right first step is 3 months of paper trading a simple rules-based system before anything more complex gets bolted on.

The recommendation

Treat a system as a tool for disciplined trading — not a money printer. The realistic outcome of doing this right is a process that:

• Removes the behavioral leaks that kill most retail traders
• Gives you a documented edge stack worth 15–25% annual at 15–25% DD
• Pays for itself once capital is €50k+
• Becomes a durable asset (the process + the track record) regardless of whether you trade it forever

The realistic outcome of doing it wrong — skipping forward-test, overfitting backtests, chasing "signal" services — is a 4-month account lifespan and a tuition payment.