Non-prediction/main-model alpha (this is HUGE), let's talk about it!
Topics:
-Regime Shift (an extension of risk mgmt)
-Position Sizing
-Trade Specific Risk Mgmt
-Whole Algo Risk Mgmt
-Execution

Starting with the risk side of things it is only natural, to begin with, a discussion on regime shift detection. I may repeat a few points from last, I may expand a few points, let's see. I'll give a brief list of all the models we can use for regime shift detection:

...

Main Models:
-HMM
-GMM
-Bayesian (regime shift, no regime shift)
-Logistic Regression (works best as an addition to all prior models, will be in the add-ons)
-Unsupervised learning (clustering by regime)
-GAN (same way they identify data properties for synthetic data...

they also suit a purpose for regime detection/classification. More broadly they are best for when you are attempting to describe regimes)
-Other NNs like SwanNet (CNN-LSTM-Wavelet which at EOD does only slightly better, but expends your fitting budget, needs work)...

Let's talk about add-ons or otherwise improvements:
-Hierarchical modelling (what timeframe did this shift occur on, short term is an alpha, long term is a risk)
-Describe the regime, not classify it, we need to know if this is a change in something we can exploit or something...

that poses a real risk to our model. A high volatility regime is a lot of times a good thing, but if autocorrelation switches and you are running a short-term crypto momentum strategy you are fucked.
-Assume T-student, obviously lol, even when using GMM TMM works better...

because with normal distributions, even if you compose a t-student distribution from normal distributions you will have a less robust distribution to changes in the tails.
-Build features, Hurst Dimension, noise filter, KPSS, loads of tests on the statistical... (more)

Short aside on some stat tests:
-Reduce Noise with a filter
-Test stationarity: ADF/VECM/KPSS
-Test dist KS-Test (or try a kernel, or decomposition like GMM/TMM)
-Non-linearity test BDS-test
-Check correlation (sum scheme/ dimensions)
-Entropy/ all that sort of good shit...

-(Max) Lyapunov exponent / time / spectrum, only max for exponent, I'm just not spelling his name 3x
-Bask Gencay Test
-Hurst Exponent/Dimension
-DFA Alpha Exponent
-Eigenportfolio features (variance of smallest/ uniformity of variance explained by them)
-Autocorrelations
...

-Spectral Embedding
-Recurrence Tests

(ok I'll stop), back to the thread:
Now we know our regime/or not, we need to work with it...

understand how that affects the probability of each trade, and whether it is worthwhile changing the PD of each trade individually wrt the observations or generally penalize all position sizing to be more conservative, may even stop if EV is now negative...

Position sizing is just kelly. What type of kelly? Well, that depends on your assumptions, but usually, half kelly is smart but even then there is a 25% chance of being in a bad position. That's the problem with discrete probability position sizing, you don't know your tails...

and thus your model probably assumes a non-fat-tailed distribution, and you have a very bad day. That is why you go continuous, or if you don't have a load of time to spend doing that, you just Monte Carlo it and use an optimization algorithm on some simulated data...

where you simulate your alphas, and then suddenly just flip them on their ass and they do the opposite so now you don't just have 0 alpha, you have -1 alpha. or more than -1 bc volatility (uh oh), that is what happens when multistrat funds sell quant btw, this is not...

just my mad ramblings (still ramblings) about tails, they exist, and you need to make sure your system is capable of knowing when it is wrong instead of just going on and on and on. It should be keeping track of the t/p vals of all these losses piling up...

and slowly become more conservative with its positioning until it finally just paper trades or otherwise records whether it would have been good/bad trade so it knows when to start being confident again, after having stopped. You need to make this a smooth transition...

which then suddenly cuts off to 0 so you don't asymptote into 20 dollar orders, logistic smooth transition, cough cough ahem ahem. Anyways, moving on to local risk/ global risk. By this, I mean pulling the plug mid-trade...

whether this is common or v. uncommon is inherent to your alpha so you need to know what sort of trades you are making. If it is momentum, losing money is negative alpha (momentum trades are known to become great mean reversion trades after finishing so GET OUT)...

if it's cyclical then you need to be a bit fancy with it, and if it's mean reversion then you stay in the damn trade until regime shift tells you alpha is no longer worth it, or the bid red button blinks and your model decides that this trade could kill it if it doesn't close...

the model for when to kill a trade that could blow up the whole strategy needs to be ENTIRELY return/EV agnostic. I DO NOT CARE how much alpha your model thinks it has on the mean-reversion trade (usually mean-reversion), it is not taking into account the risk of...

blowing up and missing out on all that future alpha. Models just aren't that smart. You also just don't want to risk a model error that makes it think it has loads of alpha when the market is blowing up. Like hey this always mean-reverts and there is loads of volatility...

and it's gone more SDs than ever, but in reality, it doesn't have huge alpha because those SDs mean nothing when your curve isn't normal, hell it isn't even a bell curve because your alpha dist looks like an M. There is no alpha for small deviatons (obv), high alpha for...

2-3SD (arbitrary, test yourself), and -alpha for anything above that bc liquidity cascades compound, regime shifts, and the presence of systemic quant risk your model won't understand. Optimize this process, but do it with a model that is transparent. Trade with NNs...

(well you don't have to, lots of my strats are NN free), but don't ever use them for risk mgmt. You should be able to describe logically why your risk patterns occur, and have a simple ensemble or tree or Markov chain for your different regime shift signals...

Remember that many of them can mean more alpha, but past a tipping point the fun volatile market with extra returns becomes dangerous and will blow you up like a balloon. Once again don't just model the market crashing, model your alphas crashing. You are probably...

not exposed to the market directly, you are exposed to the market crashing a mega funds portfolio and then they sell bc covering margin or unhappy LPs pulling out or general panick, and they are all in the same alphas as you because you really aren't special, and now you are...

becoming their exit liquidity because much like the smart beta folks you assumed your alpha was ergodic and you will always get a return. At least they have the brains to know that they are exposed to a factor which somewhat correlates to the market and not the mkt...

itself. And also don't rely on historical data for simulating alpha-factor exposures, because they would have alpha if they had already blown up (for most of them, not all, some have become their own little smart beta thing like momentum). Popular alphas don't decay...

they blow up, decay isn't enough to put them back in line, and as much as I am not an efficient markets guy, for hugely crowded alphas every pension fund runs, history has shown that the blowup usually smacks the returns right back with the rest of the pack.
-end

Not going to do execution this time. That's a whole new tweet and I have plenty on it already on this page. Go read @lightspringfox or @0xdoug if you run out of order book cage fighting ideas on my page.