CPE Research Program — Paper 16
Time-Series Momentum (TSMOM) is one of the most established, widely used trend-following strategies at systematic funds. Restricting it to only the instruments it can genuinely predict — and changing nothing else about how it trades — beats the standard, unfiltered version. The edge holds up no matter how the 20-year test is sliced.
A real, published strategy (Moskowitz, Ooi & Pedersen, 2012), implemented exactly as described — no tuning, no fitting to this program's own data.
Not a different signal — the identical TSMOM engine above, pointed at a smaller, pre-screened universe.
Same 2006–2026 history, same two strategies, re-run by splitting the period into different numbers of equal-length blocks — from 3 large blocks (~6.7 years each) down to 20 small ones (~1 year each). Each point is the average annualized return across that split's blocks.
The green line sits above the grey line at every block count tested — the gap barely moves whether the 20-year history is cut into 3 pieces or 20. That consistency, not the size of the gap, is the real result: most backtested edges narrow or flip sign under this kind of stress test. This one didn't.
TSMOM isn't a toy benchmark — it's a real strategy that systematic funds actually run, published in full detail (Moskowitz, Ooi & Pedersen, 2012) and implemented here with no tuning and no fitting to this program's own data. Improving on it usually means a new signal, a new asset class, or added complexity. Here, nothing about the strategy itself changed — only which five instruments it's allowed to trade, chosen ahead of time using a separate, already-published measurement of each instrument's own predictability structure (Ramanathan, 2026a). The improvement survives being tested ten different ways, with the historical record cut into anywhere from 3 to 20 pieces.
A narrower, secondary result in the same paper: two individual instruments (JPM, XLB) beat standard TSMOM under every real point estimate checked — using each strategy's real average-return difference rather than a regression-based alpha estimate that flips sign on over half the instruments in this program's panel. See the full preprint for that analysis and its caveats.