The Benchmark They Fined $10bn For Rigging Still Moves the Market — Legally

Once a month, at four o'clock in the afternoon London time, a large number of professional investors buy and sell currency without caring what price they get. They are not careless. They are required to. And because they all do it in the same few minutes, price moves — on the last business day of the month, pooled across nine major pairs, it moves 10.1 basis points against a normal day's 5.5.

This page measures that window on bid-side minute data: 11 instruments, four control hours, a permutation test, and one result that contradicts the version of this story most textbooks tell.

The scandal, briefly

In June 2013 Bloomberg News reported that traders at the world's largest FX banks were manipulating the benchmark rate the entire asset-management industry settles against. The mechanics were not sophisticated. Traders knew in advance which way the flow would push at 4pm, because clients had told them their orders. They shared that information with each other in private chat rooms — the regulators' published findings record the names the traders themselves gave those rooms: The Cartel, The Bandits' Club, The Mafia. Then they concentrated their own trading into the 60-second window the benchmark was calculated from, a practice with its own piece of trading-floor vocabulary: banging the close.

In November 2014 six banks were fined roughly $4.3bn between them. In May 2015 Citicorp, JPMorgan, Barclays and RBS pleaded guilty to criminal conspiracy charges, adding about $5.6bn. Across all the FX investigations, penalties ran past $10bn. The Financial Stability Board's structural remedy arrived in February 2015: widen the calculation window from 60 seconds to five minutes, so no small burst of trading could dominate it. That five-minute window — 15:57:30 to 16:02:30 London — is the one still in use.

What none of that changed is why the window was worth manipulating in the first place.

Who is buying, and why they can't shop around

Imagine a fund whose entire job is to copy an index — MSCI World, say. That index contains Japanese shares priced in yen, but the index itself is reported in dollars, so somebody has to convert. The index provider does that conversion using the 4pm London rate.

Now the fund has to match the index. If the index converted yen at the 4pm rate and the fund converts its own yen at 3pm, the fund's return will differ slightly from the index's return. That difference is called tracking error, and it is the single thing an index fund is judged on. A fund that beats its index by accident is nearly as embarrassed as one that lags it — the whole product promise is "we will return exactly what the index returns".

So here is the strange part. The fund does not want a good price. It wants the identical price. Whether 4pm turns out expensive or cheap is irrelevant, because whatever rate the index used, the fund used the same one, and the gap is zero. It is closer to being graded against an answer key than against the truth: if the key says seven, you write seven.

Two things then pile onto the last business day of the month. First, hedging. A fund holding Japanese shares normally sells yen forward to cancel the currency risk, but over the month those shares moved — so the hedge no longer matches the exposure and has to be resized. Most hedged funds and hedged share classes reset on a monthly schedule, priced at the benchmark. Second, flows: new money and redemptions are batched to month-end valuation dates. A month of accumulated drift is corrected in one afternoon, by a very large number of institutions, inside the same five minutes.

Is 16:00 London actually special?

Before testing anything about month-end, the window has to justify itself against ordinary hours. Below is the mean absolute move across the 18 minutes spanning the fix (London 15:44 to 16:02), against the identical window at three control hours, plus tick volume in the fixing minutes relative to the preceding hour.

Eleven instruments out of eleven come in above 1.0. Under a simple sign test that alone is p ≈ 0.0005, and the tick-volume column locates the cause precisely: volume peaks on the minutes 15:57 to 16:02, which is the regulated window rather than the round hour.

The metals rows are the interesting ones. Gold and silver move more at 16:00 — 1.39x and 1.27x — but their volume ratio is 0.92x and 0.99x, meaning no volume spike at all. That is the signature of an asset absorbing somebody else's flow rather than generating its own. The fix is an FX event; the metals are collateral damage.

The reversal everyone repeats does not exist on a normal day

The standard account of the fix, repeated in broker education and in a good deal of academic work, is that price drifts into the benchmark and then partly hands it back. Tested on ordinary days here, it does not. Signing each day's subsequent move against that day's run-up, only EUR/JPY clears significance (t=3.05), and every other pair sits inside |t| < 1.9. Correlation between the run-up and the following 30 minutes runs -0.06 to -0.27 — the right sign, far too weak to lean on.

Pooling properly makes it starker. Averaging the signed reversal across nine pairs and treating each date as a single observation, 350 normal days give a reversal of -0.09bp at 30 minutes (t=-0.32) and +0.01bp at 60 minutes (t=0.04). There is nothing there.

Everything is on the last business day

Splitting the same measurement by where a day sits relative to month-end changes the picture completely.

The secondary bump two days before month end was not something this study went looking for, and it is the detail that most supports the mechanism. Spot FX settles two business days forward. A trade executed on the second-to-last business day is the last one that settles inside the month, so an institution that needs its position on the books for month-end valuation has to deal by then. A settlement convention nobody trades on purpose shows up as the second-largest row in the table.

Is it month-end, or is it the fix?

Month-end could simply be a busy day everywhere, in which case 16:00 would be incidental. Running the month-end-to-normal ratio at each anchor hour separates the two.

At 11:00 and 20:00 the last business day of the month is indistinguishable from any other day. One hour after the fix it is quieter than a normal day. The p-value of 0.0007 comes from 4,000 block permutations in which the month-end labels are reassigned to randomly drawn business days — the same dates across all pairs, so cross-pair correlation is preserved under the null. The observed 2.31x sits beyond the 99th percentile of that null distribution, which tops out at 1.88x.

The honest weak point is 13:00, where month-end is itself elevated at 1.48x and the difference-in-differences only reaches p=0.083. That control window sits shortly before the 08:30 New York data slot, and month-end days carry their own scheduled releases, so it is a contaminated comparison. Against the three clean controls the fix-specific result holds at p=0.011, p=0.0012 and p=0.056.

On month-end, the move does come back

Extending the sample to five-minute bars gives 17 complete month-ends rather than six. The month-end amplification survives — median ratio 1.73x across the nine pairs — and this time the reversal appears.

Nine pairs on the same 17 dates are not nine independent experiments, so the pooled t-statistic of 3.8 flatters itself. The defensible version collapses each date to a single number — the average signed give-back across all nine pairs — and tests those 17 observations.

At 30 minutes the effect is not yet distinguishable from noise. At 60 minutes it is, modestly, and the contrast with the bottom row is the part that matters: the identical procedure on 350 ordinary days returns a number indistinguishable from zero in both columns. Month-end is not simply a louder version of a normal fix. It is the only condition under which the direction of the move carries negative information about the next hour.

Because the price data is bid-side only, a spread that widens through the fix and narrows afterwards could manufacture exactly this pattern. It would, however, do so asymmetrically — a depressed bid produces an artificial down-move followed by an artificial recovery, so the fake reversal would appear only after downward run-ups. It does not. Give-back after upward run-ups is 2.65bp (n=75, t=2.7) and after downward run-ups 2.75bp (n=77, t=2.6). That symmetry is what a real flow effect looks like, and is not what a spread artifact looks like.

What the flow does not tell you

If the mechanism is equity-hedge rebalancing, the direction ought to be forecastable: a month in which foreign equities rose leaves hedgers under-hedged in dollars and forces dollar selling at the fix. Tested against the index's own monthly return, the relationship is present in magnitude and useless in practice — correlation -0.421 (t=-1.74, n=16), a slope of -0.75bp of dollar move per 1% equity month, a jackknife range of -0.31 to -0.51, and a sign hit-rate of 8 of 16. Big equity months produce big fix moves. Which way is a coin flip.

That is the more useful finding for anyone hoping to anticipate the print. The volatility is schedulable. The direction is not.

Gold has its own four o'clock

Gold's benchmark is not the FX fix but the LBMA auction at 15:00 London, and over 25 months of minute data that hour is the larger gold event: 18.38bp against 14.15bp at 16:00 and 10.16bp at 13:00. The honest caveat is that 15:00 London is also 10:00 in New York, when ISM, JOLTS and consumer confidence print. Those two cannot be separated with price data alone, and this study does not claim to. Gold at month-end shows 1.28x against a normal day, which across 24 month-ends does not reach significance (Welch t=1.08 on log magnitude).

What to do with this

Mark the last business day of the month. It is the one scheduled distortion in FX that never appears on an economic calendar, and at 16:00 London the market moves roughly twice as far as it does on any other afternoon, on every major pair.

Do not read that move as information. On an ordinary day a sharp move usually means somebody informed is pushing, and continuation is a reasonable default. Here the buyer is an index fund executing a mandate, it does not have a view, and at 16:02 it is finished. This is the one window measured on this page where trading in the direction of the breakout has negative expectancy rather than neutral.

Do not trade the reversal either. 2.71bp will not cover a spread. The value is in the two mistakes it prevents: chasing a month-end fix spike, and leaving a stop inside a window that routinely runs twice its usual distance for reasons that have nothing to do with your thesis.

If you want to look at these afternoons yourself, the simulator replays the same bid-side archive these figures come from — load any major pair, jump to the last business day of a month and step through 15:45 to 17:00 bar by bar. Related reading: forex volatility by hour of day, what US economic releases actually do to price, and forex trading sessions explained.