Some news-trading debates come up here every few weeks. I checked a few of them against my own data. Small sample, but better than opinions.
Debate one: is it better to be flat before high-impact news or hold with a wider stop? Over about 80 releases where I had a position available to hold, holding produced a slightly better average result, but with much larger worst cases. Two slipped stops alone wiped out most of the benefit. I stay flat.
Debate two: does the first spike usually reverse? In my log of EURUSD second-tier releases, the first spike reversed at least halfway within 15 minutes a bit more than half the time. For NFP and CPI, the first move continued more often than it reversed. So it depends on the event.
Debate three: is the first 15 minutes after a release the best time to trade? For me, no. My results from 15 to 60 minutes after the release are better than in the first 15, mostly because spreads and slippage are lower.
Debate four: does trading news require a fast VPS? For trades after the first minute, my results were the same from home and from a VPS.
None of this is universal. Different pairs, brokers and sessions may give different answers. Check your own log before taking any side.
News process: common debates settled with data
-
LondonScalper
- Posts: 965
- Joined: Sat Sep 05, 2026 7:54 am
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: News process: common debates settled with data
LondonScalper, debate two is the one where I would be most careful with the sample. If the second tier log holds around 60 releases, a reversal rate a bit above half is within one standard error of a coin flip, which is roughly 6 to 7 percentage points at that size.
That does not make the finding wrong. It means I would not build a fade rule on it until the count is closer to 200, or until the edge shows up in two separate years.
Debate one fits what I found when holding through news. The average favoured holding by a small amount, but the two worst outcomes were each larger than a month of the extra gains, and that asymmetry is the whole argument for being flat.
On debate three, I wonder whether the setups are the same in both windows. My trades in the first 15 minutes were mostly breakouts, while later ones were pullbacks, so part of the difference was the setup and not only the spread.
Did the VPS comparison include manual exits, or only entries? Closing by hand in a fast market is where I noticed a gap, since each click took a second or two longer to confirm from home.
That does not make the finding wrong. It means I would not build a fade rule on it until the count is closer to 200, or until the edge shows up in two separate years.
Debate one fits what I found when holding through news. The average favoured holding by a small amount, but the two worst outcomes were each larger than a month of the extra gains, and that asymmetry is the whole argument for being flat.
On debate three, I wonder whether the setups are the same in both windows. My trades in the first 15 minutes were mostly breakouts, while later ones were pullbacks, so part of the difference was the setup and not only the spread.
Did the VPS comparison include manual exits, or only entries? Closing by hand in a fast market is where I noticed a gap, since each click took a second or two longer to confirm from home.
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: News process: common debates settled with data
LondonScalper, I went back to my own records for the second half of debate two, since that part has the clearest rule attached. Over 31 payrolls releases, the direction of the first M1 candle was still in profit 30 minutes later 19 times, which agrees with you, though the sample is about as thin as the one I questioned before.
What surprised me was how much the definition of the first move changed the answer. Using the first five minutes instead of the first candle, the figure rose to 22, because a lot of first candles were a whipsaw that reversed inside two minutes before the real move started.
So a fade rule and a follow rule can both look right on the same data, depending on where the spike is measured. I would want any rule in this thread to state its measuring window in minutes.
On debate one, I also wonder about selection. If the positions available to hold were mostly ones already in profit before the release, holding had a cushion that a fresh position would not, and the average would flatter it.
Were the 80 a mix of winning and losing positions at the moment of the release? That split alone might explain why holding came out slightly ahead.
What surprised me was how much the definition of the first move changed the answer. Using the first five minutes instead of the first candle, the figure rose to 22, because a lot of first candles were a whipsaw that reversed inside two minutes before the real move started.
So a fade rule and a follow rule can both look right on the same data, depending on where the spike is measured. I would want any rule in this thread to state its measuring window in minutes.
On debate one, I also wonder about selection. If the positions available to hold were mostly ones already in profit before the release, holding had a cushion that a fresh position would not, and the average would flatter it.
Were the 80 a mix of winning and losing positions at the moment of the release? That split alone might explain why holding came out slightly ahead.
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: News process: common debates settled with data
Some one-minute history files contain bars that were built from interpolation rather than real ticks. A missing minute gets filled by a flat or linear bar to avoid gaps in the chart.
That seems harmless but it injects false structure. A flat bar with open, high, low and close all equal tells the system that nothing happened, when the real situation might be that the data is missing. Indicators calculate on it as if it were real. A scalping signal might trigger from a series of fabricated bars.
The telltale signs are bars with identical OHLC values, or volume at exactly 1, during hours when activity should exist. Some history centers show a column for volume, and anomalies stand out.
The problem is common on illiquid pairs and on weekend or holiday data. It's also common for data converted from one format to another.
A cleanup step before testing, such as removing or flagging bars with zero range in active sessions, takes minutes. The effect on results can be larger than any parameter tweak.
A basic filter counts the number of bars in a day where open equals close equals high equals low. A normal busy day has almost none. Dozens in a row during active hours is a clear sign that the data is partly invented.
Which data source had the cleanest M1 history, and which turned out to be full of filler bars?
That seems harmless but it injects false structure. A flat bar with open, high, low and close all equal tells the system that nothing happened, when the real situation might be that the data is missing. Indicators calculate on it as if it were real. A scalping signal might trigger from a series of fabricated bars.
The telltale signs are bars with identical OHLC values, or volume at exactly 1, during hours when activity should exist. Some history centers show a column for volume, and anomalies stand out.
The problem is common on illiquid pairs and on weekend or holiday data. It's also common for data converted from one format to another.
A cleanup step before testing, such as removing or flagging bars with zero range in active sessions, takes minutes. The effect on results can be larger than any parameter tweak.
A basic filter counts the number of bars in a day where open equals close equals high equals low. A normal busy day has almost none. Dozens in a row during active hours is a clear sign that the data is partly invented.
Which data source had the cleanest M1 history, and which turned out to be full of filler bars?