A strategy library advertises breadth by counting components. The Momentum Trading Engine describes eleven distinct entry systems spanning momentum, breakout and mean-reversion, running across five synchronised timeframes, with nine strategies live at capture. Read as a count, that is a wide product. Read as a factor structure, eleven systems sorted into three families is at most three ideas, and probably fewer, because two of those three families are the same directional bet at different speeds.
The number that matters to a risk committee is not how many strategies you deployed. It is how many independent bets the book contains, and that is an empirical question with a specific answer you can compute from the trade log.
Breakout is trend with a shorter fuse
Take the taxonomy at face value first, because the structural read gets you most of the way before you touch data. Momentum and breakout both go long strength and short weakness. They differ in what triggers the entry, a persistent move versus a range violation, and in how quickly they act. Their return streams load on the same underlying condition, which is that price continues in the direction it has been moving. Mean-reversion is the genuine opposite, and it is one family, not five.
So the prior going in is that a nine-strategy book drawn from this library contains roughly two economic exposures, trend and reversion, with the split determined by how many of the nine come from each family. If seven of nine are momentum or breakout, you do not have a diversified book. You have a trend book with a small reversion hedge and nine sets of fees.
Correlate realised returns, never signals
The common error here is correlating signal series. Two systems can produce entries on different days and still deliver almost identical return streams, because what drives your P and L is exposure over time rather than the timestamp of the trigger. Correlating triggers will understate overlap badly.
What you want is a daily return series per strategy, built from the trade log. The log carries Asset, Strategy, Direction, Status, Timeframe, Order, Time in UTC, Entry, Current, PnL, Run-up, Drawdown, Source and Duration. That is enough to reconstruct a position calendar. For each strategy, mark each day as long, short or flat from the entry timestamp and the duration, apply the underlying instrument's daily return, and you have a series you can correlate.
Two adjustments before you run the matrix. Scale each series to constant volatility, otherwise the correlation is contaminated by position sizing rather than describing the signal. And be explicit about flat days. A strategy that is out of the market half the time will show a low correlation to everything simply because it has fewer non-zero observations, which reads as diversification and is really just lower exposure.
What thirty-nine trades does to a correlation estimate
The engine reports 350 trades backtested across all strategies, roughly 39 each. Before you believe any number that comes out of the matrix, price the uncertainty on it.
Using the standard transformation, the standard error on a correlation estimate over 39 observations is about 0.167 in transformed space. Work that through and a measured correlation of 0.30 carries a 95 percent interval running from roughly minus 0.02 to 0.56. A measured 0.60 runs from about 0.35 to 0.77. Those intervals overlap heavily, which means that on this sample you cannot reliably distinguish a pair that is mildly related from a pair that is substantially the same trade.
The multiplicity problem compounds it. Nine strategies produce 36 distinct pairs. Testing 36 correlations at a 5 percent threshold, you expect between one and two to clear it by chance alone. If your matrix shows two significant pairs, you have found nothing.
The honest conclusion is that at this sample length the correlation matrix cannot be used to prove independence. It can only be used to prove overlap. A high measured correlation is credible evidence that two strategies are the same bet. A low one is not evidence that they are different, and treating it as such is how books end up concentrated.
Collapsing the count into an effective number of bets
Once you have the matrix, reduce it to one number. The simple version, adequate for a risk memo, uses the average pairwise correlation across the book. The effective number of bets is the strategy count divided by one plus the average correlation times one less than the count.
Run it for nine strategies. At an average pairwise correlation of 0.60, which is what you should expect from a book dominated by two trend families, you get nine divided by 5.8, which is 1.55 effective bets. At 0.30 you get nine divided by 3.4, which is 2.65. At 0.15 you get 4.09. You need the average pairwise correlation down near 0.15 before nine strategies behave like four independent ones, and nothing in the taxonomy suggests this library gets there.
The eigenvalue version is better if you are presenting to a committee that will push back, because it does not assume a single common correlation. Take the eigenvalues of the correlation matrix, normalise them to sum to one, and compute the inverse of the sum of their squares. It gives a similar answer on a book like this and it does not collapse when a subset of strategies is tightly coupled while the rest are not.
The timeframe axis doubles the overlap
There is a second source of coupling that a naive matrix will find but nobody will interpret correctly. The engine runs five synchronised timeframes, and its own naming embeds the timeframe in the strategy identifier, with the trade log showing entries such as a Ford strategy labelled with a one day timeframe.
Two strategies on the same instrument at different timeframes share the price series exactly. Their correlation is not a statement about the entry logic at all, it is arithmetic. The daily bars are aggregations of the intraday bars, so a one day trend system and a sixty minute trend system on the same ticker are measuring the same move at two resolutions, and the shorter one will simply enter earlier and exit earlier.
Before you run the matrix, group your strategies by instrument. Any group with more than one member is a single bet on that instrument until proven otherwise, and the burden of proof sits with the strategies rather than with you. In the published configuration feed, two separate configurations ran on the same equity ticker at sixty minute and thirty minute timeframes. Deployed together those are not two positions in a book, they are one position with a staggered entry and two commission schedules.
What to do with a number like 1.6
An effective bet count near 1.6 does not mean shut it down. It means size the book as though it were roughly one and a half positions rather than nine, which changes the capital allocation by a factor of two or more against what a naive equal-weight across nine strategies would produce.
It also changes what you monitor. In a nine-bet book you watch strategies. In a 1.6-bet book you watch the regime, because every component fails on the same day. Trend books have a characteristic failure, which is a choppy range with no sustained direction, and if seven of your nine are trend, your entire drawdown will arrive in one stretch of that regime. Write that in the risk memo before it happens, name the conditions, and set the review trigger on the aggregate rather than on any individual strategy, because no individual strategy will look broken while the book is bleeding.