Kaizen FX

The problem

Forex majors are a different game from crypto perpetuals. Spreads on EUR/USD or USD/JPY are measured in fractions of a pip — the difference between profitable and unprofitable strategies often lives entirely in execution precision under those spreads. A model that backtests beautifully against mid-prices dies on first contact with the bid/ask reality. Regimes shift slowly, driven by central-bank cycles and macro flows, and play out over weeks or months rather than hours. Holding periods are days, not minutes.

These differences sound modest. They aren't. Almost every parameter that matters in a Kaizen-style trading system — feature timescales, retrain cadence, position sizing, slippage modeling, regime-detection windows — has to be re-tuned when the underlying market changes character this much. The architecture transfers cleanly; the parameters do not.

The approach

Kaizen FX uses the same model architecture as Kaizen X — XGBoost meta-ensemble, regime-aware specialist switching, learned exit timing, Thompson-sampled portfolio of variants. What changes is everything around it:

• Regime detection operates on macro and central-bank-cycle timescales rather than intraday flow.
• Slippage modeling becomes a first-class part of the signal. In crypto perps you can be approximate about fill quality; in EUR/USD a single pip of unexpected slippage flips the trade's expected value.
• Holding-period priors shift from hours to days. The exit-timing model is retrained against a fundamentally different distribution of optimal hold lengths.
• Retrain cadence slows. Nightly retrains still run, but validation windows are longer because FX moves don't generate the same density of signal per unit time.

The build

The continuous-improvement infrastructure is shared with Kaizen X: shadow deployment, drift detection, automatic rollback, the full decision-lineage log. Building Kaizen FX as a sibling to Kaizen X rather than a fresh codebase is itself a design choice — the improvement loop is the most expensive part of the system to build well, and amortizing it across both products multiplied the return on that investment.

The execution layer is the part that diverged most. Forex venues have different quoting conventions, different liquidity profiles by time of day, and different fee structures than crypto perp venues. The order-routing and fill-quality logic is FX-specific even where the model layer above it is shared.

What it does in production

Kaizen FX is live across forex majors. Same nightly self-improvement loop as Kaizen X. Same instrumentation, same decision lineage, same automatic rollback when accuracy degrades. The difference shows up in operating tempo — FX trades are slower, hold longer, react to different inputs — but the underlying discipline of “the system improves itself, by design” is identical.

The pair of Kaizen X and Kaizen FX is itself the proof that the architecture generalizes. The kaizen loop — continuous, incremental, instrumented improvement — is what Sophonix brings to every client engagement. The markets are just the domain we used to harden it.

Tech stack

• Model layer: shared with Kaizen X — XGBoost meta-ensemble + regime-aware specialist switching + learned exits + Thompson-sampled variant portfolio
• Slippage modeling: spread-aware, integrated into the signal rather than bolted on
• Regime detection: tuned for macro and central-bank-cycle timescales
• Improvement loop: shared infrastructure with Kaizen X (shadow deployment, drift detection, automatic rollback)
• Markets: foreign-exchange majors
• Execution: forex-venue-specific order routing, optimized for fill quality under tight spreads
• Infrastructure: bare-metal, shared decision-lineage logging with Kaizen X