Saturday, 1 November 2025

R squared and Sharpe Ratio

 Here's some research I did whilst writing my new book (coming next year, and aimed at relatively inexperienced traders). Imagine the scene. You're a trader who products forecasts (a scaled number which predicts future risk adjusted returns, or at least you hope it does) who wants to evaluate how good you are. After all you've read Carver, and you know you should use your expected Sharpe Ratio to determine your risk target and cost budget.

But you don't have access to cutting edge backtesting software, or even dodgy home brew backtesting software like my own psystemtrade, instead you just have Excel (substitute for your own favourite spreadsheet, god knows I certainly don't use the Micros*it product myself). You're not enough of a spreadsheet whizz to construct a backtest, but you can just about manage a linear regression. But how do we get a Sharpe Ratio from a regression?

If that is to much of a stretch for the typical reader of this blog, instead imagine that you do fancy yourself as a bit of a data scientist, and naturally you begin your research by regressing your risk adjusted returns on your forecasts to identify 'features' (I'm given to understand this is the way these people speak) before going near your backtester because you've read Lopez De Prado. 

Feels like we're watching a remake of that classic scene in Good Will Hunting doesn't it "Of course that's your contention. You're a first year data scientist. You just finished some financial economist, Lopez De Prado prob'ly, and so naturally that's what you believe until next month when you get to Rob Carver and get convinced that momentum is a risk factor. That'll last until sometime in your second year..."

But you're wondering whether an R squared of 0.05 is any good or not? Unlike the Sharpe Ratio, where you know that 1 is good, 2 is brilliant, and 3 means you are eithier the next RenTech or more likely you've overfitted.

So I thought it would be 'fun' to model the relationship between these two measures of performance. Also, like I said, it's useful for the book. Which is very much aimed at the tech novice trader rather than the data scientist, but I guess the data scientist can just get the result for free from this blogpost as they're unlikely to buy the book.

There are three ways we can do this. We can use a closed form formula, we can use random data, or we can use actual data. I'm going to do all three. Partly to verify the formula works in the real world, and partly to 

There is code here; you'll need psystemtrade to run it though.

Edit notes: I'd like to thank LacertaXG1 and Vivek Rao for reminding me that a closed form formula exists for this problem.


Closed form formula

From the book known only as G&K we have one of my favourite laws, LAM - the law of active management. This is where the famous 'Sharpe Ratio (actually Information Ratio, but we're amongst friends) is proportional to sqrt active bets' comes from, a result we use in both portfolio size space (the IDM for a portfolio of N uncorrelated assets ought to be sqrt N) and in time space (for a given success rate the SR for a trading strategy with holding period T will be sqrt 2 times better if we halve our holding period). 

Anyway under LAM at an annual holding period an R squared of 0.01 equates to an IC/SR of 0.10. Under LAM we'd expect the same R squared to result in a sqrt(256) = 16, SR of 1.6 at a daily holding period. Let's see how well this is borne out by the data.


Real data and forecasts

This is the easiest one. We're going to get some real forecasts, for things like carry, momentum. You know the sort of thing I do. If not, read some books. Or if you're a cheapskate, the rest of this blog. And we get the price of the things the forecasts are for. And because I do indeed have fancy backtesting software I can measure the SR for a given forecast/price pairing*. 

* to do this we need a way of mapping from forecast to positions, basically I just do inverse vol position scaling with my standard simple vol estimate which is roughly the last month of daily returns, and the overall forecast scaling doesn't really matter because we're not interested in the estimated coefficients of the regression just the R squared.

And because I can do import statsmodel in python, I can also do regressions. What's the regression I do? Well since forecasts are for  predicting future risk adjusted returns, I regress:

(price_t+h - price_t)/vol_estimate_t = alpha + beta * (forecast_t) + epsilon_t 

Where t is time index, and h is the forecast horizon in calendar days, which I measure simply by working out the forecast turnover (by counting the typical frequency of forecast sign changes from negative to positive in a year), and then dividing 365 by the turnover. 

Strictly speaking we should remove overlapping periods as that will inflate our R squared, but as long as we consistently don't remove overlapping periods it then our results will be fine.

Beta we don't care about as long as it's positive (it's some arbitrary scaling factor that will depend on the size of h and the forecast scaling), and alpha will be any bias in the forecast which we also don't care about. All we care about is how well the regression fits, and for that we use R squared. 

Note: We could also look at the statistical significance of the beta estimate, but that's going to depend on the length of time period we have. I'd rather look at the statistical significance of the SR estimate once we have it, so we'll leave that to one side. 

Anyway we end up with a collection of SR and the counterpart R squared for the relevant regression. Which we'll plot in a minute, but let's get random data first.


Random data

This is the slightly harder one. To help out, let's think about the regression we're going to end up running:

(price_t+h - price_t)/vol_estimate_t = alpha + beta * (forecast_t)  + epsilon_t 

And let's move some stuff around:

 (forecast_t)  

     = ((1/beta)*(price_t+h - price_t)/vol_estimate_t) 

      + (alpha/beta) + (epsilon_t/beta) 

If we assume that alpha is zero, and we're not bothered about arbitrary beta scaling, then we can see that:

 (forecast_t)  

     = ((price_t+h - price_t)/vol_estimate_t) + noise

This means we can do the following:
  • Create a random price series, compounded gaussian random is fine, and scaling doesn't matter
  • Measure it's backward looking vol estimate
  • Work out the future risk adjusted price return at any given point for some horizon, h
  • Add noise to it (as a multiple of the gaussian standard deviation)
  • Voila! As the french would say. We have a forecast! (Or nous avons une prĂ©vision! As the French would say)
We now have a price, and a forecast. So we can repeat the exercise of measuring a SR and doing a regression from which we get the R squared. And we'll get the behaviour we expect; more noise equals lower SR and a worse R squared. We can run this bad boy many times for different horizons, and also for different levels of noise.


Results

Without adoing any further, here are some nice pictures. We'll start with the fake data. Each of the points on these graphs is the mean SR and R squared from 500 random price series. The x-axis is a LOG scale for R squared. 10^-1 is 0.1 and so on, you know the drill. The y axis is the SR. No logging. The titles are the forecast horizons in business days, so 5 days is a week, etc etc.

As we're trading quickly, we get pretty decent SR even for R squared that would make you sad. An R squared of 0.01, which sounds rubbish, gives you a SR of around 0.7. 

Heres around a monthly holding period:


Two months:


Three months:


Six months:

And finally, one year:



Right so what are the conclusions? There is some fun intuition here. We can see that an R squared of 0.01 equates to a SR of 0.1 at an annual holding period as the theory suggests. It's also clear that an R squared of 0.1, which is very high for financial data, isn't going to help that much if your holding period is a year. Your SR will still only be around 0.30. Wheras if you're trading fifty times faster, around once a week, it will be around 2.30 SR with R squared of 0.1. The ratio between these two numbers (7.6) is almost exactly equal to the square root of fifty (7.1) and this is no accident; our results are in line with the law of active management which is a nice touch.

Neatly, an R squared of 1 equates exactly to a SR of 1 at a one year holding period.

Now how about some real results. Here we don't know what the forecast horizon is, instead we measure it from the forecast. This does mean we won't have neat graphs for a given horizon, but we can do each graph for a range of horizons. And we don't have to make up the forecast by reversing the regression equation, we just have forecasts already. And the price, well of course we have prices.
Important note! Unlike with fake data where we're unlikely to lose money on average, with real data we can lose money. So we remove all the negative SR before plotting.

Here's for a horizon of about 5 days:

No neat lines here; each scatter point represents an instrument and trading rule (probably mostly fast momentum). Remember this from earlier for the 5 day plot with fake data: "An R squared of 0.01, which sounds rubbish, gives you a SR of around 0.7". You can see that is still true here. And also the general shape is similar to what we'd expect; a gentle upward curve. We just have more really low SR, and (sadly!) fewer higher SR than in the fake data.

About two weeks:

About a month:

About two months:
About three months:
About six months... notice things are getting sparser
And finally about a year:
There is very little to go on here, but an R squared of 0.1 which before gave a SR of 0.3 isn't a million miles away at 0.5. In general I'd say the real results come close to confirming the fake results.


Summary

Both data scientists and neophyte traders alike can use the fake data graphs to get SR without doing a backtest. Do your regression at some forecast horizon for which a fake data graph exists. Don't remove overlapping periods. If the beta is negative then you're losing money. If the beta is positive then you can lookup the SR inferred by the R squared.

You can also use any graph, and then correct the results for LAM. For example, if you want the results for 1 day, then you can use the results for 5 days and multiply the SR by sqrt(5). But you want a closed form solution. So here is one, assuming 256 business days in a year:

The SR for N days holding period is equal to 16 * sqrt(R squared / N)



Tuesday, 7 October 2025

Is the degradation of trend following performance a cohort effect, instrument decay, or an environmental problem?

It's probably bad luck to say this, but the most recent poor performance of CTAs and trend following managers this year appears to have been reversed. My own system is up over 12% since the nadir of the summer drawdown, and is now up for year; admittedly by only by 5.5%. 

Nevetheless, it's true to say that trend following performance appears to have been degrading over the last few decades. If I can literally talk my own book (the book in question being Advanced Futures Trading Strategies - AFTS), then in one chapter I note:

Having said that, it does look like the returns from strategy nine are falling over time. The inflationary 1970s were particularly strong, with a total non-compounded return of over 500% over the decade (if we had been compounding, our returns would have been even more spectacular). We then made over 200% in each of the next three decades. But since 2010 our average return has roughly halved.

But where the better returns in the 1970's (and to a lesser extent 80's, 90's and 00's) because we had a better environment for trend following, or because we had better instruments, or because over time instrument performance decays?

Let me explain. Back in 1972 when my backtest begins, there were just a few instruments. In the first few years only 11 instruments were around, out of the 100 or so in my usual list of liquid instruments I use for testing. And they were weird: seven were agricultural commodities, two are currencies and two metals. Perhaps the better performance of trend following in the 1970's was because the instruments we had then were just better at trend following? Or, perhaps it's just that when an instrument is first traded it does very well, because there aren't many other smart people hanging around to extract 'alpha'?

So, let's see which of these three explanations is most likely.

I'm going to start with the fastest EWMAC 2,8 crossover I use. In AFTS I noted that the two fastest crossovers have suffered particularly bad degeneration in performance since about 1990. These returns are before costs, so the after cost performance would be even worse.

There are quite a few of these graphs in this format, so let me explain. Each line is a different cohort of instruments. So the blue line for example, is all the instruments that began trading in the period 1971-1980 inclusive. On the y-axis is the average SR for those instruments in the five year period beginning in the date on the x-axis. So for example, for the instruments that began trading in the first ten years, from 1981-1985 their average SR was around 0.30; a little higher than the next cohort of instruments that just came in. 

* Important note. The date an instrument starts trading is the earliest point I have data for it. It may have been trading long before then.

Let's now review three possible explanations for the reduction in trend following p&l in the fifty years or so, and see which most represents the empirical results.

First of all we have a cohort effect explanation. Instruments which enter the dataset later have worse performance, so they drag down the average performance. This would look something like this:


You can see that the older an instrument is, the better it's performance. To combat this effect we could just avoid trading newer instruments.

Next we have an instrument lifetime decay effect. Each instrument does well when it first enters the dataset, but then it's performance decays as it ages. As a result, as more instruments are old and fewer newer, and the average performance falls. An extreme version of that effect would look something like this (I've drawn the horizontal lines slightly apart for clarity, they should be on top of each other): :

To deal with this we'd have to be constantly adding new instruments that haven't yet been affected by the influx of sophisticated traders looking for alpha. This is the opposite of what we'd do with a cohort effect.

Finally we have the general enviroment effect. All instruments have roughly similar performance in each period, which gets worse over time. This would result in something a bit like this (I've drawn these lines slightly apart for clarity, they should be on top of each other): 

There is no solution here. We are buggered. We need to get out of the trend following game. Or at least hope that this is just a temporary setback; after all people have tested trend following rules over hundreds of years so a few decades of bad performance is nothing to worry about.

Looking back at the EWMAC 2,8 graph, there is no support for the cohort effect. It looks like there is some evidence of an instrument decay, most strikingly for the 1991 cohort. However this only contains 7 instruments, so it's one of the smallest cohorts, so the significance is questionable. Overall though, this does look an enviroment effect with noise. 

Let's step down the speed to EWMAC4,8:

Looks like a very similar picture. How about EWMAC8,16?
Again apart from the 1991 cohort, this looks pretty much like a gradual decline due to enviromental effects. Now let's turn to EWMAC16,64, which readers of AFTS will know is the best following of my momentum indicators:
Feels like we're watching the same film over and over again, doesn't it?

EWMAC32,128


EWMAC64,256

Not quite as clear, but we're trading very slowly now so would expect more noise.

To summarise then, it looks like the decay of momentum performance has been solely down to  general enviromental effects, rather than a cohort effect, or the decay of instrument performance. This is bad, because it means we can't do anything about it by restricting ourselves to older or newer instruments. This is good, because instrument diversification really is our best chance of being profitable traders and making the most out of a weakening signal. I wouldn't want to suggest that you should do anything different than trading all the instruments you can get your hands on.

And to reiterate, let's hope this is a temporary situation. For me personally, as someone who isn't tied into a CTA box, I'm happy to continue trying to trade as many different risk / return factors as possible.

Bonus postscript: Here's carry!

This does look a little bit more like a decaying instrument story...



Monday, 8 September 2025

PCA analysis of Futures returns for fun and profit, part deux

 In my previous post I discussed what would happen if you did the crazy thing of doing a PCA on the whole universe of futures across assets, rather than just within US equities or bonds like The Man would want you to. In this post I explore how we could do something useful with them. There is some messy code here, to run all of it you'll need psystemtrade, but you can exploit big chunks with your own data even if you don't.


The big problem: sign flipping

Before hitting some p&l generating activity, first however we need to deal with an outstanding issue from the previous post.

TLDR, most of the time factor one is 'risk on /equities are go' and factor two is global interest rates; although not always. Factor sign flipping was a problem however (thanks to people below the line for that insight). So sometimes factor one was long equities, sometimes short equities. Sometimes it was something else entirely. 

As an example, remember this plot from part one? It's the factor exposure of the S&P 500 over time for factors 1(0), 2 (1) and 3(2).

Note there are 'blips' when we have a short exposure to factor 1, mostly in the period since 2008 when we're normally long factor 1. That's clearly a temporary sign flip. We probably want to get rid of those. But there is also the long period in the early 2000's when we're persistently short factor 1. That might be a 'sign flip'; but it could also be that factor one in this period was something more interesting than just 'long equity risk'. 

A couple of ideas spring to mind here. One is just smoothing the factor weights. That would easily solve the blips; and the smooth need only be a few weeks to get rid of them. But a longer smooth, of the length needed to get rid of the other periods, would reduce the information about the factors; in particularly we'd be missing out on interesting times when something other than boring old risk on and off is driving the market.

Another bright idea I had was to reverse the sign on weights when the largest absolute value weight was negative. My expectation was that generally the largest weight on factor 1 (mostly risk on) would usually be equities, and when that factor flipped sign we'd flip it back again. However that didn't produce the expected results. If I were less lazy (and eager to get back to writing book #5), I'd probably do some research; eg I'm pretty sure the answer is somewhere in Gappy's new book but I haven't got there yet. 

In the end I decided to relax and ignore the sign flipping; I can do this because of the four ideas I outlined:

1- own the factors

2- trade the factors

3- buy assets with persistent alpha (+ve residual) 

4- mean revert the cumulative residual 

.... it's only really 1 and 2 that are affected by sign flipping. And I feel I already have things in my armoury for 1 and 2. For example my aggregate momentum signal (blogged about here, and also in my most recent book AFTS) is basically like 2, and on assets with a long bias that will also give us a chunk of 1 as well. 

<Sidebar * note to browser not actually HTML>

Arguably my relative momentum and long term mean reversion are also a bit like 3 and 4. Yet another idea is to build 'asset classes' using clustering as I did here, and then use those for the purposes of 1,2 and possibly 3 and 4. 

So we have three different ways of forming 'factors': exogenously determine asset classes, PCA, and clustering; and four different ways of trading each of them. Those won't give radically different results since clusters mostly follow asset classes, but they could be a little different.

<\Sidebar * see previous note>

But, I hear you cry, why can you flippantly ignore sign flipping when trading only the residuals? Well it's pretty simple; consider a standard APT type equation with a single PCA k and market i:

r_i_t = a_i + (b_i,k * r_k,t) +e_i,t

If we now do a sign flip, then the beta (b) will have a minus one in front of, but the market or PCA return r_m will also have a minus one in front of it. These cancel, and estimation of both the persistent bias (alpha, a_i) and the temporary error (epsilon, e_m) will be unaffected. 


Trading the alpha

So we have two basic ideas; we generate our PCA and then run regressions that look like this:

r_i_t = a_i + (b_i,k * r_k,t) + ... +e_i,t

Where there are one or more PCA k.... And then we eithier buy positive a_i and sell negative; or we sell things with recent cumulative positive e_i.

There are still many design questions to resolve here. How many PCA do we include? Too few, and we'll probably end up missing something interesting. Too many and there is a risk we'll end up without clear signals. Over what period should we estimate betas and alphas? Basically how persistent are they likely to be. Over what period should we cumulate epsilon? Are there periods in which episilon will be trending rather than mean reverting; eg assets that have outperformed their factor adjusted return will continue to do so (which will look an awful lot like buying positive alpha)?

For the PCA I'm going to keep it simple and initially use three PCA, which happens to be the most I can plot and get my head around it. I'm also going to stick to estimating my alphas and betas over a 12 month period, which is the arbitrary period I used before to estimate the PCA themselves (seems weird to use a different period). For the question of epsilon decay I will risk the wrath of the overfitting gods and do a time sensitivity analysis.

To summarise then: At the start of each month we look at the 12 months normalised returns, do a PCA, and then regress each instrument on the returns of each component. We then have an alpha intercept coefficient, and some betas (at most three, once for each PCA). We can see how predictable the alpha is of returns in the following month(s). Then for the following month we can also calculate the residual of performance vs the fitted model. We can cumulate up these residuals and see how they forecast performance.


Alpha

Let's start with the alphas. Here be a massive scatter plot:


Each point is the alpha calculated at the start of a given month for an instrument, and the normalised ex-post return for the following month. It looks like there might be a weak positive relationship there, so let's do some stats.

                           OLS Regression Results                            
==============================================================================
Dep. Variable:         ex_post_return   R-squared:                       0.006
Model:                            OLS   Adj. R-squared:                  0.006
Method:                 Least Squares   F-statistic:                     169.0
Date:                Mon, 08 Sep 2025   Prob (F-statistic):           1.62e-38
Time:                        11:46:09   Log-Likelihood:                 1910.6
No. Observations:               26311   AIC:                            -3817.
Df Residuals:                   26309   BIC:                            -3801.
Df Model:                           1                                         
Covariance Type:            nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
Intercept      0.0052      0.001      3.735      0.000       0.002       0.008
alpha          0.3144      0.024     12.999      0.000       0.267       0.362
==============================================================================

There: we have the classic undergrad stats exercise question "Why is my t-stat big but my R squared is low?". Answer: there is something here, but it's weak. This often happens if you use a large dataset (26,000 observations here).

To show this differently, the conditional ex-post average daily return over the following month with a positive alpha is 0.02, and with a negative alpha is -0.01 (both conditional subsets are roughly half the dataset overall). The t-statistic comparing these is a hefty 10, corresponding to a p_value of the order of 10^-26. So again, alpha definitely has an effect, but is that difference really that big? Hard to tell.

But we know that low R squared are pretty common in finance, so is this a problem? To test this I tried using the alphas as a forecast, and then calculated the Sharpe Ratio of each forecast. The median across all instruments is a SR of 0.1. Remember that trend following gives us around 0.3 to 0.4 for each instrument, so this isn't especially interesting.

It might be that I would get better results from a different lookback to calculate the alphas (remember we use one year). Everything from a 1 month to a lookback of 10 years. What about using fewer, or more principal components? Remember we're going with three.  It turns out that a one year lookback is pretty optimal compared to shorter or longer; but using one PC is better than using two or more. Still the very best we can do is a one year lookback with one PC, and that gives us a SR of 0.12 which is hardly in wallet busting territory and also not significantly different from the result wih three factors.


Trading the residual

Let's turn then to trading the residuals. We're going to cumulate up residuals over various periods and see how well that predicts future returns. To avoid a forward looking forecast, the residuals are calculated on the out of sample month following the point at which the model is fitted. Otherwise the regression coefficients would be forward looking, and hence so would the forecast.

Note that because the model changes slightly each month, the coefficients used to calculate residuals will also change slightly. Such is life. But we'll keep stacking up the residuals month by month even though they are using different models.

We now have 3 knobs to twiddle on our overfitting machine; lookback and number of PCs as before, but also the number of days we sum up residuals. To keep things relatively simple I will initially sum them up using 22 days (about a month of business days). So our base case is:
  • One year lookback to do PCA and calculate coefficients
  • 3 principal components
  • 22 days summing up of residuals; our forecast is minus the summed residuals
And jumping straight to Sharpe Ratio calculation like an impatient toddler, we get a SR of -0.04. The sign is wrong, showing that positive residuals lead to more positive performance, and the effect is also v.v.v. weak.

Does increasing the residual summing period work; eg mean reversion works over longer time periods? Nope. Anything up to a year is actually worse. Going down to a week (which would be v.v.v. costly to trade) does at least push the SR into positive territory, but only just.

Dropping to one PC (which was marginally better than for alpha above), changing the lookback on the PCA, .... nothing produces useful results. This idea is an already dead donkey that has been subsequently thrown off a cliff and then burned**.

** no actual donkeys were harmed in the creation of this blog post

 

Back to alpha

So we had a not too promising individual SR using alpha on an instrument level, but how does that look on a portfolio level? Surprisingly, quite good. Here are the combined results for 100 instruments:


That bad boy has a SR of 0.84! Some of that is diversification, but some of it is because a more accurately calculated median SR per instrument is 0.15, higher than the 0.10 calculated earlier (because, many reasons, like buffering and what not). Still, that's an extremely high realised diversification of over 5. Let's compare it to the 'gold standard' of single momentum models, EWMAC16,64 with the same instruments:


OK not as good, even counting the different vol ewmac comes in with a SR of 1.08, but their correlation is a relatively lowly 0.3. That suggests a modest allocation to alpha persistence will earn some money. Chucking 10% of your forecast weights into alpha persistence bumps up the SR of ewmac16,64 from 1.08 to 1.12. Going to the (arguably in sample fitted) best model with one principal component improves the SR of the alpha model by itself to 1.02, but also increases the correlation with momentum; so that the joint SR with 10% in alpha persistence and 90% in ewmac produces a pretty much unchanged SR of 1.13.


Summary

Research in systematic trading tends to result in a lot of blind alleys. I thought this would be another one. Certainly the idea of mean reverting the errors, a classic from the equity stat arb crowd, doesn't really work in this context. However there does seem to be some modest performance gain to basic momentum from including a PCA derived alpha persistence model. The gain is small however, so it's debatable whether it's worth what would be quite a lot of additional work. Not a blind alley then, but not a very pleasant one to spend much time in.