Build and test a strategy Lesson 4 of 14
Strategies · Building and testing a strategy
How to read a backtest report, line by line
Read a backtest report on a real test of QQQ: what total return, CAGR, deepest fall, time invested and win rate each say, and what each one hides.
On this page
A backtest report is a short table and a chart. It looks like an answer, and most people read it as one: they find the return line, compare it with the market, and stop. The lines underneath, and the things the table leaves out, usually matter more.
This lesson reads one real report from top to bottom. The test is the 10-month moving average rule from Moving average trend following: the 10-month rule, tested, run on QQQ, the largest Nasdaq-100 fund, from January 2000 to October 2026. It was picked because it is awkward. The rule earned almost exactly what holding the fund earned, lost money on most of its trades, and still did something holding could not. Every line of the report is needed to see all three facts at once.
The report in front of us
The rule: at each month-end, hold QQQ if its close is above the average of its last ten month-end closes; otherwise hold cash. The conditions of the test:
- Data. QQQ month-end closes from April 1999 to October 2026. Ten months are needed for the first average, so the test runs from January 2000: 322 months, 26.8 years.
- Trading. The signal and the trade use the same month-end close.
- Costs. 0.10% of the amount on every buy and every sell.
- Left out. Dividends, interest on cash and taxes.
| Measure | 10-month rule | Buy and hold |
|---|---|---|
| Total return | 651.4% | 669.9% |
| Return per year (CAGR) | 7.8% | 7.9% |
| Deepest fall from a peak | −36.9% | −81.1% |
| Longest time below a peak | 90 months | 181 months |
| Time invested | 71% | 100% |
| Round-trip trades | 27 | 1 |
| Trades with a gain | 48% | n/a |
This is the standard table on every strategy page in this library. Start above it, with the four bullets. A report without them cannot be read at all, because every number below depends on them. If a published backtest does not say what it charged for trading or which prices it traded at, the right assumption is the generous one: no cost, and a price nobody could have got.
Reading the table, line by line
Total return
One dollar became $7.51 under the rule and $7.70 held throughout. Total return is the most quoted number and the least useful for comparing tests, because it grows with the length of the test. A 650% return over 27 years and a 650% return over 10 years are different animals. Use it only to compare two lines of the same report.
Return per year (CAGR)
The compound annual growth rate turns the total into a steady yearly rate: 7.8% for the rule, 7.9% for holding. It can be compared across tests of different lengths. What it hides is the path. No year of this test looked like 7.8%. A CAGR is an average of a journey that, in this case, included a fall of four-fifths.
It also inherits the test's dates. More on that below, because for this report it is the most important caveat of all.
Deepest fall from a peak
This is the maximum drawdown: the largest drop from any high to a later low. Holding QQQ lost 81.1% from March 2000 to September 2002. The rule lost 36.9% from March 2000 to February 2003.
Read this line as a question about the reader, not the strategy: would this fall have been sat through? Someone who sells at the bottom does not earn the CAGR on the line above. A deepest fall of 81% means the 7.9% a year was available only to a person who watched $10,000 become $1,900 and did nothing.
Two cautions. The deepest fall in a backtest is the worst of the past, and the future may hold a worse one. And the number here uses month-end closes, so a fall that was deeper in the middle of a month is not seen.
Longest time below a peak
The rule spent 90 months, seven and a half years, below its March 2000 value, and first passed it in October 2007. Holding spent 181 months below, and did not see its March 2000 value again until May 2015.
This line measures patience, where the one above measures pain. Many strategies are abandoned during a long flat stretch and not during a crash. A report that shows only the depth of the worst fall, and not its length, leaves out the part that tests people most.
Time invested
The rule held the fund in 71% of the months and cash in the rest. Two things follow. The rule earned its 7.8% while exposed to the market less than three-quarters of the time, which is the sense in which it took less risk. And the idle 29% matters to the left-out list: cash earned nothing in this test, where a real account would have earned interest, and the rule missed the dividends of the months it sat out.
Trades
Twenty-seven round trips in 26.8 years, about one a year. The count tells how much the cost assumption matters and how much evidence there is. With 54 or so transactions, raising the cost from 0.10% to 0.50% per trade lowers the yearly return from 7.8% to 7.0%; with no cost at all it is 8.0%. A rule that traded every week would be far more sensitive.
Twenty-seven is also a small sample. Conclusions about the rule's average trade rest on 27 observations, which leads to the last line.
Trades with a gain
Forty-eight percent: 13 winners, 14 losers. Read alone, this line says the rule is a coin flip. Read with the sizes, it says something else. The average winner gained 25.5% and the average loser lost 5.0%. A strategy can be right less than half the time and make money, or be right 80% of the time and lose it. The win rate is only meaningful beside the two averages, and the standard table on these pages always gives them in the text for that reason.
The trade list: what the averages hide
The table summarizes. The trade list is the evidence itself, and it is worth looking at every row.
Three things stand out that no line of the table shows.
The result rests on a few trades. The mean trade gained 9.7%, but the median trade lost 1.1%. The typical trade lost a little; a handful won a great deal. The best trade, April 2020 to January 2022, gained 65.5%, and the second, January 2023 to March 2025, gained 58.8%. Take the best one out and the final value falls from $7.51 to $4.54. Take out the best two and it is $2.86. A report whose result depends on two trades out of 27 is fragile in a specific way: it depends on the person being in the market for those two, having sat through the fourteen losses that came between the winners.
The losses cluster. The first three trades all lost: 7.5%, 4.5% and 12.8% between January and September 2000. From May 2004 to May 2006 four trades in a row lost. A person who began this rule in 2000 started with three losses and then spent more than two years in cash.
The worst trade is modest. The largest single loss was 12.8%. The deepest fall of 36.9% was not one bad trade. It was a run of small ones during a falling market. Reports that give only "largest losing trade" understate what a strategy can lose.
Four questions the table does not answer
Does it depend on the dates?
Split the period and the story changes. This is the most useful single check on any report.
| Period | 10-month rule | Buy and hold |
|---|---|---|
| January 2000 – February 2009 | −0.6% a year | −12.2% a year |
| February 2009 – October 2026 | 12.5% a year | 20.0% a year |
The near-tie over the whole period is two large gaps in opposite directions. Up to the February 2009 low the rule lost 5.6% in total while holding lost 69.3%. Since then the rule has earned 12.5% a year and holding 20.0%.
The starting date does a lot of work here. The test begins in January 2000 only because the fund began trading in 1999, and that happens to be two months before the top of the dot-com bubble. Start the comparison at the September 2002 low instead and the rule earned 9.9% a year to October 2026 against 15.7% for holding. A reader shown only one of those two tables would come away with opposite opinions of the same rule. Neither table is wrong. The honest report shows both.
Does it depend on the setting?
The rule has one number in it, the length of the average. Changing it by two months either way gives:
| Average length | Return per year | Deepest fall | Round trips |
|---|---|---|---|
| 6 months | 8.9% | −50.6% | 30 |
| 8 months | 6.4% | −57.9% | 31 |
| 10 months | 7.8% | −36.9% | 27 |
| 12 months | 7.8% | −26.2% | 22 |
Each test starts when its own first average exists, so the first months differ slightly. On SPY the same four lengths gave similar results to each other. On QQQ they do not: the yearly return runs from 6.4% to 8.9% and the deepest fall from 26.2% to 57.9%. That spread is a warning. With so few trades, moving one buy or one sell in 2000 or 2008 by a month changes the whole result, so no single line of this table deserves much trust. Picking the 12-month line because it has the smallest fall would be overfitting: choosing the luckiest of four tries.
What would it have cost in a real account?
The test charges 0.10% per trade and nothing else. In a US taxable account, each closed winning trade would have produced a taxable gain in the year it closed, several of them held less than a year. Holding defers all of that. On the other side, the cash would have earned interest and holding would have collected more dividends. A report cannot settle these for every reader; it can only list them, and a report that lists nothing has not thought about them.
Is one fund enough?
No. This is one instrument over one stretch of history that happened to contain the worst crash a major US index fund has had. The same rule on SPY from 1993 returned 7.8% a year against 8.9% for holding, a clearer loss on return. A result that appears on one symbol only is an observation about that symbol.
A reading order that works
- Conditions first. Data, dates, trade price, cost, what is left out. If any is missing, stop.
- Compare like with like. Find the benchmark line. A return with nothing beside it cannot be judged.
- Fall before return. Read the deepest fall and the time below a peak, and ask whether they could have been lived through.
- Win rate with sizes. Never one without the other two.
- Count the trades. Fewer than about 30 means every average is rough.
- Open the trade list. Look for the two or three trades that made the result, and for clusters of losses.
- Split the period and move the settings. A result that survives both is worth more than a better result that does not.
How Finabase does it
Finabase has two testing tools, described in Backtest indicator signals and rule-based strategies on Finabase. Their reports differ from the table on this page, and knowing how avoids misreading them.
- Backtest trades every signal of the studies chosen on one symbol. Each signal buys or shorts a fixed $10,000, with no trading cost. Its report gives Gain in dollars, Gain / max cost (the gain as a percentage of the most capital tied up at once), Trades with wins and losses, Win rate, Signals in range, a market value and cost chart, and the full transactions table. It has no deepest-fall line and no buy-and-hold line, so compare the gain with the symbol's own change over the same range on the chart, and read the transactions table the way the trade list is read above.
- Strategy replays if/then allocation rules over two years from a starting amount and shows Return, Account value, cash, market value and the trades. Two years holds two or three signals of a monthly rule. That is enough to see how a rule behaves and far too little to judge it.
Because the Backtest tool sizes every signal at the same dollar amount, its gain is not a compounded return and cannot be set beside the CAGR in this lesson. Neither tool tells anyone what to buy or sell.
Frequently asked questions
What is the most important number in a backtest report?
There is no single one, but the deepest fall from a peak is the number most often skipped and most often decisive. It says what holding the strategy would have felt like at its worst, which decides whether the return on the line above it could really have been earned.
What is a good win rate for a trading strategy?
There is no good or bad win rate on its own. The rule in this lesson won 48% of its trades and multiplied the starting amount by 7.5, because its average winner was five times its average loser. A strategy that wins 80% of the time loses money if its rare losses are large enough.
How many trades does a backtest need?
More than most reports have. With 27 trades, as here, removing two changes the outcome completely. As a rough guide, averages built on fewer than about 30 trades are loose, and a rule that trades once a year cannot produce many more than that from the history that exists.
Why does the same strategy show different results on different sites?
Because the conditions differ: the dates, whether dividends are counted, the cost per trade, and whether trades happen at the signal bar's close or the next bar's open. Two honest tests of one rule can differ by a percentage point a year for those reasons alone.
Summary
A backtest report is read from the conditions down. The return lines say how the test ended; the deepest fall and the time below a peak say what it was like on the way; time invested, the trade count and the win rate, read with the average win and loss, say how the result was produced. In the QQQ example the rule matched buy and hold on return, 7.8% against 7.9% a year, and had less than half the worst fall, but that summary depends on a start two months before a historic peak, on two trades out of 27, and on a setting whose neighbours gave quite different numbers. Splitting the period, changing the setting and opening the trade list are what turn a report from a claim into evidence.
Check your understanding
4 questions on this lesson. Each answer comes with the reason, and nothing is scored.
A report shows a 7.8% yearly return and nothing about trading costs or which prices the trades used. What is the sound way to read it?
- As a 7.8% return, since costs are small
- As unreadable until the conditions are known; assume no cost and an unobtainable price
- As a return somewhat above 7.8%, since tests are usually cautious
Show the answer
B. As unreadable until the conditions are known; assume no cost and an unobtainable price Every number in a report depends on its conditions: the data, the dates, the trade price, the cost and what is left out. A report that omits them has usually made the generous choice on each.
The QQQ rule won 48% of its trades and still turned $1 into $7.51. How?
- Its average winner gained 25.5% and its average loser lost 5.0%
- It traded very often, so small edges added up
- Trading costs were left out of the test
Show the answer
A. Its average winner gained 25.5% and its average loser lost 5.0% A win rate says nothing without the size of the wins and losses. Thirteen winners averaging 25.5% outweigh fourteen losers averaging 5.0%. The test charged 0.10% per trade, and the rule traded only about once a year.
Without its two best trades, the rule’s final value falls from $7.51 to $2.86. What does that tell a reader?
- The two trades were errors in the data
- The rule is certain to fail
- The result depends on being invested for a few large moves, having sat through the many small losses between them
Show the answer
C. The result depends on being invested for a few large moves, having sat through the many small losses between them The mean trade gained 9.7% and the median trade lost 1.1%: most trades lost a little and a few made nearly everything. Anyone who stopped following the rule after a run of losses would have missed the trades that carry the result.
Over the whole test the rule and buy and hold tie at about 7.8% a year. Split at February 2009 they are −0.6% against −12.2%, then 12.5% against 20.0%. What is the lesson?
- A whole-period figure can hide two opposite results, so a report should be split at least once
- The rule stopped working in 2009 and should be ignored
- The first half is the true result because it contains a crash
Show the answer
A. A whole-period figure can hide two opposite results, so a report should be split at least once The tie is an average of a period that suited the rule and one that did not, and the test happens to begin two months before a historic peak. Neither half is the true one. Showing both is what makes the report honest.
Exercise
One task to do in Finabase. Nothing is graded or saved.
Open the sample backtest and scroll to the Transactions table. Read down the Realized gain column and pick out the two largest gains. Add them and compare the sum with the Gain at the top of the page. Then look for losses that come one after another. How much of the result rests on two trades, and do the losses cluster?
Open the sample backtestWhat you should find
Judge it the way the lesson reads its trade list. If the two best trades make up a large share of the Gain, the result is fragile: it depends on having been in the market for exactly those two. Losses that arrive in a run are what a person following the signals would have had to sit through. The summary lines show neither, and the page has no deepest-fall line and no buy-and-hold line.
Sources and further reading
- Mebane T. Faber, “A Quantitative Approach to Tactical Asset Allocation,” Journal of Wealth Management (Spring 2007) — the 10-month rule used as the example
- David H. Bailey, Jonathan M. Borwein, Marcos López de Prado and Qiji Jim Zhu, “Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance,” Notices of the American Mathematical Society 61(5), 2014
- SEC Office of Investor Education and Advocacy, “Investor Bulletin: Performance Claims” — on hypothetical and back-tested results
- Finabase chart data — QQQ monthly bars, April 1999 to October 2026; rule and results computed by Finabase
