Robert Carver spent more than seven years at AHL, one of the world’s largest systematic hedge funds, finishing up as the manager of a multi-billion dollar fixed income portfolio. When he left in 2013 he spent his time writing a book, Systematic Trading, and building a fully automated system to trade his own money. He’s been public about both since then, which makes him a rare case: someone who has seen systematic trading from the institutional side and can translate it clearly for everyone else.
The episode covers two things in depth. First, what actually makes a trading rule good, and how most traders get this wrong. Second, a detailed look at walk-forward testing, why the standard application of it is much weaker than most people assume, and what a more robust alternative looks like.
Watch the full episode below, then read on for the complete breakdown.
What actually makes a good trading rule
Robert’s definition of a systematic trading rule is precise: it needs to be something you can write down algorithmically, test on historical data, and that someone else (or a computer) could replicate without subjective interpretation. Traders who look at a chart, see a pattern, and say “I use trading rules” aren’t being fully systematic unless they can codify exactly what they’re doing.
Beyond testability, his criteria for a good rule are:
- Not overfitted. The rule should work across a variety of market conditions, not just the specific period you tested it on.
- Relatively simple. More parameters mean more ways to fit past data without capturing something real. “More complicated rules are more likely to have been overfitted and relatively simple rules, less so.”
- Understandable. If you can’t explain what the rule is doing and why, you have a black box. A black box can’t tell you when the conditions that made it work have changed.
- Has an underlying explanation. The rule should reflect something that makes intuitive sense about how markets behave, not just a pattern that happened to appear in the data.
A trend following rule is the example he keeps returning to because it passes all these tests clearly. If price has been going up, it tends to carry on going up. That’s a simple idea, it’s understandable, and there’s a plausible economic explanation for why it persists. You can identify conditions under which it would stop working. You can’t do that with a complex black-box model.
Robert’s view on alpha: there isn’t much of it
This is where Robert’s perspective cuts against a common belief in systematic trading. He doesn’t think there’s a lot of hidden alpha to be discovered. “I don’t really believe there are a huge number of trading rules that exist, which are doing something extraordinarily interesting and going to make you lots of money.”
What he does believe exists are persistent effects, what some economists call “alternative beta.” Things like the equity risk premium (equities outperform bonds because they’re riskier), the term structure premium (longer-dated bonds yield more), and trend following (momentum works across asset classes over long periods). These aren’t secrets. They can’t be called alpha because they don’t require special skill. But they’re real and they’re exploitable.
His portfolio is built around these kinds of well-documented, persistent effects rather than unique or novel signals. “I personally like to put into my portfolio things that aren’t necessarily a surprise to people and are kind of very well-known effects.” The bet is that these effects have plausible explanations and will persist, rather than that he’s found something nobody else has spotted.
The problem with walk-forward testing
Walk-forward testing (or rolling window backtesting) is widely used and widely recommended. The idea is that you avoid look-ahead bias by only using data from a past window to fit your model, then testing on the period immediately after. Robert’s critique isn’t that the approach is wrong in principle. It’s that the window lengths most traders use are far too short to give you statistically meaningful results.
“A year is a very, very short period of time to actually find any sort of statistical evidence about whether something is working or not.” If you’re fitting on one year of data and forward testing on the next, you’re essentially seeing whether a pattern that appeared in one market regime persists into the next, and one-year windows aren’t large enough to separate signal from noise.
His research shows you need at least ten years to make a reasonable statistical judgment about whether a rule is profitable, and ideally longer to compare one rule against another. With a ten-year window requirement and a typical backtest starting in 1980, you can’t even begin doing true walk-forward testing until around 1990. That’s not much out-of-sample data left to work with.
“So in practice the walk-forwardness becomes a bit of a moot point because if you imagine you say, well, I really need 20 years of data to make a good statistical judgment and you start your backtest in 1980, well, for the first 20 years of that back test you’re not going to be moving forward.”
The finding that surprises most traders: averaging beats selecting
Here’s where Robert’s research gets genuinely counterintuitive. The standard walk-forward approach is to take a set of similar trading rules (say, moving average crossovers with different lookback parameters), identify which one performed best in the past window, and use that one going forward. His tests show this approach is significantly outperformed by a different method: just take an average of all the rule variations.
“Trading rule returns are pretty noisy from year to year or even over longer periods. And it’s really hard to pick out which one is the best. And by far the best approach is just to take an average.”
The intuition behind this is related to overfitting. The rule that happened to perform best in the last five-year window was probably at least partly lucky. That luck doesn’t necessarily persist into the next window. By averaging across multiple parameter variations, you reduce the impact of any single rule’s idiosyncratic performance and get a more stable signal.
The practical caveat: you don’t want to average across rules that are highly correlated with each other. If you have dozens of moving average crossovers with slightly different parameters, many of them will correlate above 99%. Averaging those adds almost no benefit. What matters is selecting rules that are reasonably distinct (correlation below 95% or so), spaced across the parameter space, and then averaging those.
Continuous rules vs. binary rules
Most trading rules are implemented as binary signals: long or flat, long or short. Robert’s book argues for a different approach, using continuous signals that scale position size proportionally to the strength of the signal rather than switching on and off.
The advantage is smoother portfolio behavior. A binary trend rule creates sudden position changes when price crosses a threshold. A continuous rule adjusts position size gradually as the signal strengthens or weakens. This reduces transaction costs (you’re not creating full-size trades on marginal signals) and smooths the equity curve.
It also has a statistical benefit. Rather than losing all the information in a signal by reducing it to a binary yes/no, you preserve the gradient. A strong trend signal might produce a larger position than a weak one. Over many trades, this tends to improve the signal-to-noise ratio of the overall strategy.
How Robert built his own system after leaving AHL
After seven years of institutional portfolio management, Robert found himself building a retail-scale version of what he’d been doing professionally. The constraints were different. Smaller capital means fewer instruments and higher transaction costs as a percentage of the portfolio. But the principles were the same.
He implemented the system fully automated, which is important because it removes discretion from execution. He trades his own money, which means he doesn’t have the investor reporting and regulatory obligations that came with the AHL role. That simplification let him focus entirely on the trading logic itself.
His book, Systematic Trading, documents the full framework he uses, including position sizing, rule selection, and portfolio construction. He’s been transparent about the methodology because he genuinely believes there isn’t much danger in revealing it. The rules he uses aren’t secrets. They’re well-documented effects. The implementation discipline is what most people struggle with, not the ideas themselves.
What institutional traders do that retail traders don’t
Robert’s time at AHL gave him exposure to practices that don’t appear in most retail trading education. The most useful of these relate to how you think about trading rules:
- Institutional desks run many variations of core strategies simultaneously, not a single optimized rule.
- They evaluate strategies over much longer periods than most retail backtests cover.
- They prefer robust rules with clear explanations over complex models that fit past data well.
- They treat the parameter selection problem as a diversification problem, not an optimization problem.
The retail equivalent of these practices is accessible. You don’t need AHL’s infrastructure. But you do need to let go of the idea that there’s a single best parameter set waiting to be discovered, and that walk-forward testing on short windows can reliably find it.
Get the show notes & transcript
Related episodes
- Andreas Clenow on trend following in stocks and institutional position sizing
- Jerry Parker on 30+ years of trend following and why simple rules last
- David Aronson on testing indicators and evidence-based technical analysis
- Nick Radge on what separates robust trading systems from fragile ones
Want to learn more about building better trading systems? Subscribe to the Better System Trader podcast for weekly interviews with the world’s top systematic traders and quantitative researchers.
