← All articles

Round numbers as support and resistance: eight instruments, eight failures

11 min read · Published 28 Oct 2026

Traders like round numbers: they place their take-profit orders on them more often, and their stops just beyond them. Studies of real orders at a currency-dealing bank from around the turn of the century showed that this is not a myth. We checked whether you can make money from it today, on gold, six indices and USD/JPY in 2018–2023. A limit order at the nearest round number in the direction of the trend lost money after costs on all eight instruments (PF 0.83–0.93), and after the correction for eight tests, round numbers turned out no better than random levels at the same distance.

Abbreviations: a limit order is a pending order that buys below the current price (or sells above it) and is filled only when price reaches it. 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; 1.00 means breaking even. ATR is the average candle range, a measure of volatility. H1 and M5 are hourly and five-minute candles. Placebo is the same procedure on levels that are not round.

Hypothesis

A round number does not come from price history. Nobody derives it from highs and lows; it is simply easy to remember and to type. That is why so many people place orders at it. This sets the idea apart from ordinary support and resistance drawn on a chart, which failed several times in our earlier tests.

There is concrete research behind the idea. In a 2000 paper published by the Federal Reserve Bank of New York, Carol Osler showed that support and resistance levels published by six firms identified the points where intraday trends ended on the currency market better than random levels did. In a 2003 paper (Journal of Finance) she looked at the orders themselves: she analyzed almost 9,700 orders from a large currency-dealing bank from 1999–2000 on three pairs. Take-profit orders clustered exactly at round numbers, and stop orders just beyond them. This gives the textbook picture: price reverses more often at a round number, and once it breaks through, it accelerates, because other traders’ stops are triggered.

Our hypothesis, frozen before the first run, was this: a limit order in the direction of the trend, at a round number no more than 2 ATR(H1) from price, gives a PF of at least 1.25 after costs. It should also do better than the same mechanism at a random level the same distance away, and better than a market entry at the moment of decision.

We wrote down our expectations before the start, and they were modest. We put the chance of passing on even one instrument at 20–25%. The effect had been measured on orders from the late 1990s, before algorithms came to dominate trading, and it might have disappeared since.

How we tested it

Definition of a round number. One for all instruments, with no tuning: a multiple of a “three significant figures” step. On gold above $1,000 this gives a level every $10 (1,850, 1,860…). On the broad US index in the test years the step was 10 points, and on indices quoted in tens of thousands (the German and Japanese indices and the index of 30 large US companies) it was 100 points. On USD/JPY it comes out as a whole yen (110.00, 111.00), exactly the “round figure” from Osler’s papers. Separately, we calculated levels ten times sparser and “rounder”, for example 1,900 and 2,000 on gold or 13,000 and 14,000 on the German index.

The mechanics were shared with two other level tests (chapters 14 and 15), so that only the source of the levels changed:

• eight decision points a day at fixed times, Monday to Friday;
• direction from the EMA50 on H1: above the average, buys only; below it, sells only;
• up to three round numbers in the direction of the trade, no further than 2 ATR(H1), the deepest ones within reach;
• a stop 1 ATR(H1) beyond the level, so usually outside the zone where, according to Osler, other traders’ stops sit;
• the position split into two halves with targets of 1 R and 2 R; once the trade is 1 R in profit, the second half gets its stop moved to the entry price;
• an order lives for 240 minutes, at most three decisions per day, and fills and exits are calculated on M5 candles.

Instruments and period. Gold, the broad US index, the US technology index, the index of 30 large US companies, the German, Japanese and UK indices, and USD/JPY. The years 2018–2023 were the training sample, and the period from 2024 to September 2026 a one-shot holdout sample, only for instruments that passed training.

Cost was the median spread from ticks over the last 30 days, plus slippage. In units of risk it ranged from 0.9% on USD/JPY to 8.6% on the UK index. As it later turned out, on gold and USD/JPY it was understated, because tick history does not include the account markup (more in the lesson at the end). On the broad US index and on the UK index the cost was the highest relative to volatility, which we recorded as a warning before the start.

Placebo and criteria. The placebo shifted each level by 0.25–0.75 ATR, so that it lay no closer than 0.1 ATR to any round number, while keeping the distance from price (200 draws). The benchmark was a market entry at the moment of decision, with the same stop and position management. Passing required everything at once: at least 100 trades, a PF of at least 1.25, a PF without the three best trades of at least 1.10, an edge over the placebo and in the permutation that survives the correction for eight tests (the Benjamini–Hochberg procedure), at least 60% of years in profit, and a result better than the market entry. A clock check (chapter 2) was built into the script: gold’s volatility peak fell at 13:30 UTC, where it should.

Result

Training sample 2018–2023, average result per decision after costs:

No instrument came close to the PF 1.25 threshold, so we did not open the holdout sample. All eight times the verdict was KILL.

On three instruments (gold and the Japanese and UK indices), round numbers did marginally better than random levels: placebo p between 0.035 and 0.05. With eight tests at once, however, the smallest p had to fall below 0.00625, and none came close. Even where round numbers “beat” the placebo, they still lost money. The limit did worse than a plain market entry on seven instruments out of eight. The exception was the UK index, where the market entry was the most costly in the whole table.

Did “rounder” levels work better?

The sparser levels did better on four instruments and worse on four. The best result, PF 1.01 on the German index, means 179 trades and zero profit. There is no sign that the “rounder” the number, the better it worked.

Why it doesn’t work

The trace of information is smaller than the cost. Round numbers are not entirely random: on three instruments they beat random levels at the edge of significance. The previous day’s levels on gold gave the same picture (chapter 15). A level carries a little information, but too little to cover the spread and slippage. If you added the cost back, the result per decision would be roughly zero, from about −0.05 to +0.01 R.

A limit collects bad fills. An order at a level in the direction of the trend is filled when price has moved against the trend all the way to the level, and it misses the situations where price went our way straight away. In chapter 14 we show this effect on three sources of levels at once: there, the limit lost to the market entry in 9 comparisons out of 11, and 7 of those 9 were round numbers.

The direction alone was empty too. A market entry in the direction of the average gave a PF of 0.83 to 1.01. So there was no edge for a round number to improve.

Orders at a currency bank are not our market. Osler’s papers described the currency market at the turn of the century, where people entered orders by hand. Today algorithms know the same literature and can front-run clusters of orders at round numbers. On top of that, orders sit where the main trading takes place. The gold futures contract is usually quoted a few to several tens of dollars away from the spot price, so “2,000” on one market is not “2,000” on the other. This is our explanation, not a measurement: we did not check on which market the orders that could “hold” price were sitting.

For the order grid on gold, this was the second and last attempt. After the previous day’s levels, round numbers and a profile of real exchange volume (chapter 13), all calculated on the same day with the same engine, we closed the topic of placing limits at levels for good. We no longer test any further source of levels for limit orders.

The lesson for you

A round number is a place where people put orders, but it is not an entry signal. In our data, a limit entry at a round number in the direction of the trend gave roughly zero before costs and a loss after costs, on every one of the eight markets.

How to check this yourself:

1. Compare the level with a placebo. Shift each of your “important” levels by half an ATR in a random direction and run the same strategy. If the result is similar, it is not the level doing the work.
2. Compare the limit with a market entry at the same moments, with the same stop and target in R. Without that, you don’t know whether waiting for the level gives you anything.
3. Calculate cost as a percentage of risk. With a cost of 8% of R, a strategy has to earn more than 0.08 R per trade on average before costs just to break even.
4. Use one definition for all instruments. If you choose the instrument or the level step after seeing the result, you will measure your own luck, not round numbers.
5. Check the cost on real trades. The spread from tick history does not include the markup our account adds: on gold and USD/JPY our cost in this test was understated (chapter 3). So the result would have been even worse.

Osler’s research also suggests a second, more cautious use: don’t put your own stop just beyond a round number, because that, according to this research, is where other traders’ stops cluster. This we did not test.

In your trading journal, mark whether the entry was at a round number, and compare those trades with the rest. The “Luck or edge?” calculator will show whether the difference goes beyond chance.

Educational content, not investment advice. Past results do not guarantee future results. We describe the result of our reconstruction of a single, strictly defined mechanism of limit orders at round numbers, on our data and with our costs — not an assessment of every way of using round numbers.