← All articles

Better data, stronger dollar? Economic news has a direction for two minutes

10 min read · Published 27 Oct 2026

The textbook says: if US data come in better than forecast, the dollar strengthens and gold gets cheaper. We tested this with criteria written down before the first calculation, and found no direction. Then we found two of our own mistakes in that test. A correct measurement showed where the direction really is: in the first two minutes, out of reach for an ordinary trader. This chapter tells both stories, because the second teaches more than the first.

R is a result in units of risk (−1 R = a loss equal to the distance to the stop). PF (profit factor) is total gains divided by total losses. ATR is the average range of a candle, here an hourly one. Placebo is the same procedure run at moments when the cause under study was absent.

Hypothesis

A surprise is the difference between the published figure and the market’s forecast. Stated at its best, the hypothesis is this: data stronger than forecast mean a stronger economy, higher interest rates and a stronger dollar. So gold should fall, USD/JPY should rise, and — in the mapping we adopted — the German index should fall too. The data are hard, public and available to everyone in the same second. If trading “on the news” has a basis anywhere, it is here.

We also had a reason of our own. We wanted our assistant, built on a language model, to read news reports and set a direction based on them. Before asking whether the model gets it right, we wanted to know whether there was anything to get right at all. If the sign of the surprise on its own does not predict direction, a model reading the same numbers — with a delay and the risk of misreading them — has nothing to discover.

We froze the mapping “strong data → gold down” in advance. We also wrote down that a significantly opposite result would not count as success, because reversing the rule after seeing the result is curve-fitting, not an edge.

How we tested it

Source. We used the economic calendar built into the trading terminal. It was the only one that had both historical actuals and forecasts; popular online calendars do not even include a field for the actual figure in their export. The export covered January 1 – September 17, 2026: 9,902 events, including 366 top-importance USD releases, of which 302 had both an actual figure and a forecast.

An event is one release minute. Several readings in the same minute were counted together, and the surprise was taken as the sign of the average of their signs. We excluded: minutes in which the signs cancelled out (19); events with a Fed speech, the release of Fed minutes or a Fed meeting in the window from −15 to +30 minutes (7), because there is no number to compare there; and the event with a major release for another currency within ±5 minutes (1). That left 151 events. The oldest 90 (from January 2 to June 3, 2026) formed the training sample. We set the newest 40% aside as a one-shot holdout sample, to be used only if the training sample passed the criteria.

Measurement. The reference point was the close of the one-minute candle just before the release, and the outcome was the close 30 minutes later. The main instrument was gold, with USD/JPY and the German index as replications.

Controls frozen before any counting: - placebo: the same times of day on days without major US data, five draws per event; - sign permutation in daily blocks, 2,000 draws; - correction for three simultaneous tests (the Benjamini–Hochberg procedure); - the result without the three events with the largest move and without the weekly crude oil inventories, because those say nothing about the state of the economy; - the share of months with a hit rate above 50%; - a separate move-size condition: the median 30-minute move must exceed twice the transaction cost measured from ticks.

Result

After the correction for three tests, no result was significant. In line with the decision table we had written down, we closed the topic, and the holdout sample stayed untouched. On gold, the mapping hit less often than the placebo: stronger data lifted gold more often than they pushed it down (p = 0.086, not significant). We had written down in advance that we would not turn such a result into a “the other way round” rule.

The move-size condition passed with an enormous margin. The median gold move in the 30 minutes after a release was $10.97, about 0.50 of the hourly ATR, against a cost of $0.09. We later measured that the real cost on gold, including the account markup, is about $0.19. The move is still dozens of times larger. There is a move after the data, and it is large. What it lacks is a direction you could predict from the sign of the surprise.

What later research showed

The question came back later in two other measurements (we describe them in more detail in the following chapters).

Reaction map 2018–2022. This was a description, not a test with criteria. It covered 1,125 release minutes, with the calendar time converted correctly and with the sign reversed where a higher number means a weaker economy (unemployment, jobless claims). The result: in the first two minutes the direction is consistent. With a clear surprise, the mapping “strong data → gold down” got the direction of the jump right in 63% of cases. But the jump in gold is explained mainly by the move in USD/JPY in the same minute (R² 0.60), not by the published number itself (R² 0.03): gold reacts to how the interest-rate market reads the data. After two minutes nothing is left: the correlation of the jump with the move over the next 28 and 58 minutes is +0.04 and +0.05.

Scalping the breakout in the first minute. We tested a rule: enter in the direction of the breakout when two one-minute candles close outside a ±1 ATR band, with a short stop and a 2 R target. On 803 releases from 2017–2022 (gold and bitcoin), the result was −0.18 R per trade and a PF of 0.72. Even at zero cost it would have been only a PF of 1.15, and the breakout direction was no better than a randomly chosen direction (p 0.24).

Why it doesn’t work

The forecast is already in the price, and algorithms price the surprise within seconds. The market knows the expectations before the release. The only new information is the gap between the actual figure and the forecast, and that gets translated into rates and the dollar before a human has time to read the headline. By the time you can realistically get in, the directional move is already done.

One minute, several conflicting numbers. The jobs report gives employment, the unemployment rate and wages all at once, and they can say different things. The sign of one number does not answer the question of how the market will read the whole.

The state of the market sets the strength of the reaction. When gold moves together with bond yields, it reacts strongly to data. When it doesn’t, the reaction almost disappears. So the same release produces a large effect one time and none the next.

What is left for the trader is volatility. A wide move in both directions, widened spreads and slippage are a risk, not an opportunity. That is why our system blocks new entries around the most important releases, and the measurement in this chapter is the quantitative justification for that block.

An honest postscript: what was wrong with our test

While writing this chapter, we reread the script from September and found two mistakes. Then we checked them against the data.

1. Calendar time and chart time drifted apart in winter. The terminal’s calendar records time in a fixed time zone, while the chart uses the broker’s time zone, which shifts with US daylight saving time. In our file, the jobs report sits at 16:30 in January, February and early March, and at 15:30 from April onwards. In reality it is always released at the same time. So until March 8, 2026 we were measuring a window that began an hour after the release. The clock check before the test covered only summer months and did not catch this.
2. We did not reverse the sign for “more is worse” data. For jobless claims and the unemployment rate, a figure above the forecast means a weaker economy, but the test counted it as strong data.

In the training sample, the window was shifted by an hour for 39 of the 90 events, the sign was wrong for 19, and 11 had both errors at once. Only 43 of the 90 were measured correctly. So the September test is not proof that there is “no” direction. A hit rate below one half and a placebo that beats the signal are exactly what you would expect when more than half of the measurements are looking in the wrong place. The later reaction map removed both problems and reached a more precise conclusion: there is a direction, but it lasts about two minutes. We are not recomputing the old test on the same data, because that would be a second shot at a sample we have already seen.

The lesson for you

News almost always moves the market. The first question to answer is whether the move that is left after your real delay lets you extract a direction after costs.

How to check this yourself:

1. Measure the move on your instrument from the moment you can realistically enter. If you need 30 seconds to decide, the first 30 seconds are of no use to you.
2. Write down the direction rule before collecting the data, and compare it with a placebo: the same times on days without a release.
3. Check the sign and the clock. Unemployment and jobless claims have the opposite sign. The US jobs report always comes out at 8:30 a.m. New York time. Convert that to your chart’s time zone and check a few dates in winter and in summer. If it “moves” by an hour in your calendar data, the calendar and the chart are in different time zones.
4. Calculate the cost at the moment of the release, not in a quiet hour, because that is when spreads and slippage are at their widest.
5. Treat major releases as a risk. Decide in advance what you do with open positions and whether you open new ones at all.

In your trading journal, tag trades opened within 30 minutes of major data and evaluate them separately. The “Luck or edge?” calculator will show whether the difference between them and the rest goes beyond chance.

Educational content, not investment advice. Past results do not guarantee future results. We describe the results of our own measurements, on our data and with our costs.