Thread by @macrocephalopod

Couple of thoughs about trend following is probably the most straightforward systematic trading strategy to implement - you're just buying the stuff that's been going up and selling the stuff that's been going down.

It's big business though - the CTA industry (which is largely trend followers) manages ~$350bn in funds and managed accounts, and there is probably half that again in in-house implementations, multi strategy funds etc.

I've implemented very simple and very complex trend following strategies, with assets ranging from a few tens of millions to a few billions, so thought it might be interesting to talk about how they work and how you would go about building one.

I'm not going to get into a discussion about whether markets trend, or what might cause trends. From my point of view it is obvious that trends exist and can be exploited systematically. This thread is about how to build a trend following strategy *assuming that trends exist*

The strategy has a few pretty well defined components -

1. Universe selection
2. Trend detection (aka signal)
3. Mapping from trend strength to desired position
4. Sector/asset class weights
5. Portfolio risk targeting
6. Trading rules
7. Execution

Universe selection. You want to trade as many *diversifying* markets as you can. The performance of trend following on a single market is pretty bad, around 0.2 Sharpe. But the strategy scales like sqrt(breadth), where breadth is the number of independent markets you trade.

For example, if you already trade S&P 500 and Nasdaq futures, then adding DJIA futures to the strategy will give almost no benefit. But adding something uncorrelated (e.g. natural gas futures) will help a lot.

Most trend followers trade futures, or other derivatives that allow them to use leverage. A typical futures trend follower might trade 60-80 futures markets on the major exchanges (CME, ICE, Eurex). A bigger and more sophisticated fund might trade 300-400 markets, including many

unusual or niche futures markets (e.g. power, cryptocurrencies, freight) as well as FX forwards or NDFs, interest rate swaps, or CDS indices. I don't think it's unrealistic to say you can increase the Sharpe ratio of the strategy by 1.5-2x by expanding the trading universe.

The trend signal is obviously important, but perhaps less than you think. It's normally quite obvious whether a market is trending up or down. The challenge is to capture this quantitatively. The most common trend signals are variations on moving average crossovers - for example,

a fast moving average of (log) price minus a slow moving average, probably normalized by volatility. These have the advantage of being simple and hard to overfit, whilst also filtering out the higher frequency return noise (they act like a low pass filter).

It's typical for trend followers to try and detect trends on horizons from 1 month to 1 year, with an average around 3-6 months. Some managers specialise in short-term trend following, even intraday, but they are generally much smaller. It's rare to focus on >1y trends.

Once you have the trend signal, it's common to pass it through another function which maps the signal to a target risk allocation for that market. Normally you don't want to continue building the position as the trend gets stronger - you want to cap out at some point.

Two common approaches are a sigmoid function (left) or a function like x * exp(-x^2) which reduces the target position for very large values of the trend signal (sometimes called "overextension")

Going to take a break but I'll continue with sector weights, portfolio risk targeting and trading/execution later.

Obviously I had a typo right in the first tweet, FML.

Picking this up. One thing you might do here is add a bias to this function for markets which are expected to go up on average (eg equities, credit or fixed income). This tends to improve Sharpe though I personally don’t like doing it because it weakens one of the nice

properties of the strategy — its ability to profit in fixed income or equity bear markets. However empirically it seems like a lot of trend following managers do this.

Once you have this mapping from trend strength to target exposure, the most common way to size positions is to make them inversely proportional to a market-specific volatility forecast. The idea is that for a given trend strength, each market contributes equally to risk,

to maximise diversification. Volatility forecasts can be as complicated as you like, but you get most of the benefit even with very simple forecasts, eg a blend of a steady-state long term volatility (estimated over 10 years) and a short-term realized volatility (eg 30-60 days)

Sector-specific weights are something you can do to try and get some extra juice out of diversification. For example, if you trade equity indices and fixed income, but have 2x the number of equity futures as you do rates futures, then equal weighting all markets leaves you

overweight equities. So you might increase the size of your your rates positions by a factor of 2 to balance this. Or maybe you trade both rates and commodities, with the same number of markets in each, but correlations between commodity markets are lower than between rates mkts.

Without any adjustment, commodities would contribute less to your portfolio risk than rates, so you might size up the commodity positions until each sector contributes equally to risk in the long run.

There are many approaches to this, from simple equal weighting to complex hierarchical or pca-based methods. My current favorite is to bucket up the universe into broad asset class groups, and try and have roughly equal volatility from each asset class in the long run.

Portfolio risk targeting is about controlling overall risk of the strategy. If you don’t do anything here you will run very little risk when not many markets are trending, and a lot of risk when there are many strong trends. The range can be like 0.3x to 3x your target risk.

If you’re happy to sometimes run a 60% vol portfolio when your target is 20% maybe that’s fine, but lots of people don’t like it (especially if you are running the strategy on behalf of a conservative LP, like a pension fund or investment consultant).

Something you can do here is size up positions a bit when there are few trends, and size them down when there are many trends, to try and stay closer to your target risk level. That sounds crazy but empirically it seems to improve sharpe ratios.

You end up with a more consistent PnL stream, at the cost of reducing your potential gains when trends are very strong (which empirically seems to happen in big equity or rates bear markets). The right answer here probably depends on why you are allocating to the strategy and

how it fits in with the rest of your portfolio (eg is it an alpha source alongside many other uncorrelated strategies, or are you hoping to use it as a partial hedge for long-only equity or fixed income positions you hold elsewhere).

Taking another pause here, more later.

We’re back.

Trading rule. You don’t want to trade on every change in your signal or vol forecast. You would just churn volume and cost yourself money in commissions/slippage. Better approach is to use buffering, as described in the post below.

https://twitter.com/macrocephalopod/status/1373216917058682881

This helps you stay pretty close to your ideal target positions, while reducing turnover and therefore reducing transaction costs. You can vary the size of the buffer with market conditions, eg making it wider when spreads are wide, and for low vol assets (where your

positions and trades will be larger, so impact of churn is higher). If you are running a large book (say $100m or more) you might also want to only trade partially toward the target, so that you get into positions over multiple days, but that’s not usually necessary for

a smaller book trading liquid markets.

The advanced version is to use a portfolio optimiser, though for a simple strategy like trend it’s not really necessary (and it creates more ways for things to go wrong, so may be better to skip it and use simpler heuristics instead)

Finally, execution. The big choices here are whether you do the execution in house or delegate to a broker, and whether you use an algo or get a human trader to work it. None of this is particularly different for trend following than any other slow strategy so I won’t get into

it much other than to say that algos (including broker algos) are fine for almost all liquid markets, but for something less liquid or more niche there are still situations where a human trader can add value.

I think that’s it. If you have particular questions put them under this tweet and I’ll try to answer as many as I can.