Fx Markets

FX markets

FX breaks three assumptions an equity toolchain makes: there is no consolidated tape, the quote convention is not uniform, and a position earns or pays interest every night.

1. 🚨 A spot-only backtest is missing the carry

A currency position is two interest rates. Holding long AUD/short USD earns the AUD rate and pays the USD rate every night — the swap or rollover. That is not a fee; for a carry trade it is the entire thesis.

✅ Demonstrated in scripts/fx_conventions.py: a long AUDUSD position over the sample returns spot only −2.97%/yr (Sharpe −0.27) — a loser — and total +1.34%/yr (Sharpe +0.19) — a winner, on the same price series. The sign of the result flips.

If your FX backtest computes pct_change() on a spot series and stops, it is not a backtest of a position anyone can hold. Add carry_return(spot, r_base, r_quote, days).

✅ Covered interest parity verified live to 0.001% across two CME contracts. The size of the omission: leaving FX rollover out overstates a long-EURUSD backtest by 1.30%/yr — or 3.91%/yr of equity at 3x leverage, which is where most FX strategies actually run.

⚠️ Retail rollover is not the interbank differential — brokers mark it up, often asymmetrically, so the carry you actually receive is smaller than parity implies and the carry you pay is larger. Model the broker's published swap rates, not the policy rates, when the strategy is carry-dependent.

2. 🚨 Pip size is not uniform — and getting it wrong is a 100× error

JPY pairs quote to 2 decimal places (pip = 0.01). Almost everything else quotes to 4 (pip = 0.0001).

✅ Demonstrated: sizing a position for a fixed pip risk on USDJPY with the 4-decimal assumption produces a notional of $15,025,000 instead of $150,250 — 100× oversized, turning a budgeted $1,000 loss into $100,000.

This is the FX analogue of the Greeks scaling error in ../../../fin-core/skills/derivatives-pricing/SKILL.md §4: an arithmetic mistake that raises no exception and is invisible in the number itself.

Note also pipettes — most venues now quote a fifth decimal (third for JPY), so "the last digit" is a tenth of a pip, not a pip.

3. Quote conventions

EURUSD means units of USD per 1 EUR — EUR is the base. The market convention is not "always USD first":

A long EURUSD is long EUR and short USD. A long USDJPY is long USD and short JPY. Treating the pair symbol as "the asset" and going long it means opposite USD exposure depending on which side of that table you are on — and the sign error survives every plot.

Crosses (EURGBP, AUDJPY) are usually synthesized from two USD legs, so their spreads are wider than the majors and their tick data is often reconstructed rather than observed.

4. 🚨 There is no consolidated tape and no official close

FX is over-the-counter. Every venue has its own price, and "the" closing price does not exist. Consequences that break research:

• Two data sources will disagree, and neither is wrong. Reconciling them is not a data-quality task.
• Daily bars depend on an arbitrary cut — 17:00 New York is the most common convention, but a vendor using 00:00 UTC produces different daily returns from the same underlying market.
• The WM/Refinitiv 16:00 London fix is the closest thing to an official benchmark, and it is a fixing window, not a print. If your strategy trades "the close", say which close.
• Backtests that assume a single global price cannot model the venue selection a real execution would face.

4b. 🚨 Free FX bars are quantised, and futures invert the convention

✅ Yahoo's hourly EURUSD is quantised to 1.34 pips — 6.7x the real 0.20-pip spread. Measured: 44 distinct values across 116 bars, all exactly representable in float32, with 1/p landing on a 1e-4 grid. There is no bid/ask and volume is all zeros. A spread or microstructure study on it is measuring the storage format, not the market.

🚨 CME FX futures invert the spot convention for JPY, CAD and CHF. 6J quotes 0.00645 while USDJPY quotes 156.10 — reciprocals. Joining a futures series to a spot series without inverting produces a perfectly plausible, entirely wrong correlation.

✅ DukasCopy free tick FX is real — verified end-to-end: 8,717 ticks in one hour with genuine bid/ask and a median spread of 0.20 pips. Two gotchas: the month in the URL is zero-indexed (January is 00), and findatapy's base URL now 301-redirects.

5. Value dates and settlement

Spot is T+2 for most pairs, with T+1 for USDCAD (and a few others). That matters for carry accrual — the rollover is charged on the value date, so a Wednesday position typically accrues three days of swap to cover the weekend.

NDFs (non-deliverable forwards) exist for restricted currencies — KRW, TWD, INR, BRL, CNY offshore conventions differ from CNH. A "USDCNY" series may be onshore, offshore, or an NDF, and those are different instruments with different prices.

6. Data

🔑 Free tick-level FX exists: DukasCopy via findatapy (freq='tick', fields=['bid','ask']) is the only free tick source in this catalogue — see ../../../fin-core/skills/market-data-sourcing/references/findatapy.md.

For point-in-time interest rates to compute carry honestly, use FRED/ALFRED vintages (../../../fin-core/skills/fundamental-and-macro-data/references/fredapi.md) — policy rates are revised and republished, and using today's rate history to compute yesterday's carry is the same look-ahead as any other macro series.

🚨 forex-python is alive but serves ECB reference rates, and the ECB itself states: "Using the rates for transaction purposes is strongly discouraged." They are a daily 16:00 CET fixing for accounting, not a tradeable price. Backtesting execution on them is not a backtest.

🔴 Dead FX/broker clients: v20, fxcmpy, forexconnect, oandapyV20 have all rotted. ib_insync is archived — use ib_async.

⚠️ Prefer a real vendor plus your own conventions layer over a wrapper that hides which side of §3 you are on.

7. Scripts

scripts/fx_conventions.py — parse_pair, is_inverted, pip_size, pip_value, notional_for_pip_risk, carry_return, total_return. Its demo is the spot-vs-total sign flip and the JPY 100× sizing error.