ETF Flows Chase the Trend - Leveraged Products Fade It Instead

Porting a systematic-trading trend/breakout forecast to ETF NAVs, and testing it against future flows and price returns across roughly 800 mainstream and leveraged/inverse products.

A strong recent trend predicts mild price mean‑reversion over the following one to two months - in both mainstream and leveraged ETFs. But the money moves in opposite directions relative to that fading trend, depending on which one you're holding.

We built a continuous trend-strength score for ETF NAVs - a risk-adjusted blend of exponential moving-average crossovers and rolling breakout channels, in the style used by systematic trend-following funds - and tested whether it predicts what happens next to a product's flows, NAV return, and AUM, at four horizons (5, 10, 20, and 40 trading days). Across 189 mainstream “Pareto” ETFs and 604 leveraged/inverse ETFs with enough history, a strong recent trend predicts mild price mean-reversion over the following one to two months in both groups. But the two groups' money moves in opposite directions relative to that trend: Pareto flows keep modestly chasing a trend that is already fading (median r = +0.09 at 40 days), while leveraged/inverse flows increasingly do the reverse - buying dips and selling rallies - and that contrarian tendency strengthens with leverage, from a coin flip at 1x to a 71%-majority pattern at 3x.

Why we looked at this

Our earlier work on flow persistence asked whether daily ETF flows have “memory” - whether an inflow tends to be followed by more inflows. That told us something about the shape of flows in isolation, but not whether flows have any relationship to the thing investors are presumably reacting to: the product's own price trend.

A natural way to test that is to build an explicit, continuous measure of how strong and how established a product's current price trend is, and see whether it leads, lags, or has no relationship with flows. Rather than invent one from scratch, we adapted a well-established one from systematic trading: Robert Carver's forecast-scaling framework combines an exponentially-weighted moving-average crossover - the difference between a fast and a slow EMA of price, risk-adjusted by the instrument's own annualised volatility - with a rolling breakout-channel measure, into a single score conventionally capped at plus or minus 20.1 We ported that construction from its usual home (futures price series) to daily ETF NAV history, then tested it against three things: daily flow as a percentage of prior AUM, daily NAV return, and daily AUM growth, at multiple forward and backward horizons.

We ran this across two of our existing dashboard universes: “Pareto” (roughly 200 mainstream megacap ETFs - iShares, Vanguard, SPDR, and similar) and “Leveraged/Inverse” (roughly 1,200 products, 2x/3x and short funds), so we could see whether a relationship looks the same for a buy-and-hold index fund as for a product built for tactical, short-term use.

What we found

Two directions matter here, and they mean very different things. Looking backward - does the past 5–40 days of NAV return or AUM growth line up with today's trend score - is close to circular, since the score is built from trailing price history; it strengthens from around 0.1 at 5 days to around 0.8 at 40 days in both universes, and mainly confirms the score measures what it's supposed to, not a discovery about investor behaviour. The forward-looking direction - does today's score predict what happens over the next 5–40 days - isn't guaranteed by construction, and it's where the two universes part ways.

Price tends to mean-revert after a strong trend - more so in leveraged/inverse

But flows move in opposite directions relative to that fading trend

Figure 1. Pareto's flow correlation is small but consistently positive and grows with the window - money keeps modestly following a trend that's already starting to fade. Leveraged/inverse's flow correlation starts negative at short horizons and only turns mildly positive by 40 days - a materially different, more front-loaded-contrarian shape.

AUM growth (not shown separately here) sits between the two at every horizon in both universes, exactly as it should given AUM growth is arithmetically close to NAV return plus flow.

Why leveraged/inverse differs: it's about leverage, not direction

The aggregate “leveraged/inverse flows are contrarian” number turned out to be hiding a cleaner and more specific pattern once we split it out by leverage tier. We looked at the correlation between the trailing 40 days of cumulative flow and today's trend score for each product - a negative number here means recent outflows have coincided with a currently strong uptrend (or recent inflows with a currently strong downtrend): classic “sell the rally, buy the dip.”

At 1x - simple, unleveraged inverse hedges, the ProShares Short S&P 500 type of product - that pattern is a coin flip: roughly a third of products lean contrarian, a third lean trend-chasing, a third show no clear relationship either way. Once leverage steps up to 2x, a clear majority (59%) turn contrarian. At 3x, that majority grows to 71%, with trend-chasing down to just 9% of products.

Figure 2. As leverage increases, a coin flip becomes a clear contrarian majority. The gradient held whether the product was long or short the underlying at a given tier - leverage magnitude is the driver, not bull-versus-bear positioning.

Two real products illustrate the two ends of this pattern clearly. We checked both against raw NAV/shares data first, the same discipline our earlier flow-persistence work required after one candidate example turned out to be an issuer-wide split artifact rather than genuine trading - neither SOXL nor SH showed any split-boundary contamination or repeating-oscillation pattern in the analysis window.

SOXL · Direxion Daily Semiconductor Bull 3x. Shows the contrarian pattern about as cleanly as it gets (r = −0.85 at 40 days).Trailing 40-day inflows reached as high as +90% of AUM during the 2025 chop and decline, while the trend score sat around -10, an established downtrend. As NAV then rallied from roughly $90 to $300 into mid-2026 with the score holding near +10, trailing 40-day flow moved to roughly -70% of AUM, the inverse pattern playing out in the other direction.

SH · ProShares Short S&P 500. Shows the opposite (r = +0.72 at 40 days) - though it's one of the strongest examples within its own tier's evenly-split population, not a typical one. Inflows cluster exactly when SH's own price is rallying hardest, which is when the hedge is paying off. That's tactical hedgers adding to a position that's working - genuine trend-following, not profit-taking.

Both patterns are consistent with existing literature on this corner of the market. Leveraged ETFs are well known to exhibit stronger path-dependent volatility drag than their 1x counterparts purely from daily rebalancing mechanics,2 consistent with the stronger price mean-reversion we found in leveraged/inverse products. Separately, recent work on market-wide levered-ETF rebalancing finds that contrarian investor flows - buying the dip - measurably offset the rebalancing demand that would otherwise follow a big index move, while momentum-chasing flows amplify it.3 Our finding is consistent with a good chunk of the higher-leverage, retail-facing universe sitting on the contrarian side of that split.

Takeaways

• A price trend fading is a broadly reliable pattern, more so in leveraged/inverse products. Strong recent trends predict mild mean-reversion over the following one to two months in both universes, and more strongly in leveraged/inverse (median r = −0.29 vs. −0.18 at 40 days).
• Mainstream and leveraged/inverse money move in opposite directions relative to that trend. Pareto flows keep modestly chasing a trend that's already turning; leveraged/inverse flows, especially at 2x and 3x, do close to the opposite - buying weakness and selling strength.
• This is a leverage-magnitude effect, not a bull/bear-direction effect. The contrarian share of products rises steadily with leverage tier (1x: a coin flip, 2x: 59% contrarian, 3x: 71% contrarian) regardless of whether the product is long or short the underlying.
• Treat any single-number “flows lead/lag price” claim with caution. The backward-looking, by-construction correlations here are large (up to 0.8) and mean very little; the genuinely informative numbers are the smaller forward-looking ones.

Next step: the same question, segmented by issuer and by underlying asset class (single stock vs. sector vs. broad index leveraged products), to see whether the leverage-tier gradient we found is itself driven by a narrower set of especially popular, retail-heavy single-name products (SOXL, CONL, and similar) rather than leveraged/inverse products generally.

Methods, for the curious

The trend/breakout score, in plain terms

For each ETF, we calculate three “fast-EMA-minus-slow-EMA” trend measures (EMA pairs of roughly 2–4 weeks vs. 3–12 months), each divided by the ETF's own trailing annualised volatility so a quiet bond fund and a volatile single-stock leveraged fund are judged on a comparable scale, then blend those three speeds with three rolling breakout-channel measures (is today's price near the top or bottom of its recent 20/40/80-day range) into one number, conventionally capped at plus or minus 20. A positive score means an established, risk-adjusted uptrend; a negative score, an established downtrend; values near zero mean no clear trend either way. This is the standard forecast construction from systematic trend-following,1 applied here to ETF NAV history instead of futures prices.

A methodological fix worth flagging

ETF NAV data isn't automatically split-adjusted the way most price feeds are. Feeding raw NAV into this kind of trend/volatility calculation means a real stock split - a clean, order-of-magnitude NAV change with no economic content - gets read as a giant single-day “return,” which corrupts the volatility estimate and the score for weeks afterward. We caught this on a real case (two State Street sector SPDRs, XLE and XLU, both 2:1 split on 2025-12-08) before running the full analysis, and fixed it by feeding the score a split-adjusted NAV series built from this project's existing split-detection pipeline - the same one used for flow calculations - rather than the raw figure.

Why two directions, and why the backward one is nearly circular

“Does the past 40 days of NAV return correlate with today's score” is close to tautological because the score is a function of trailing price history by construction - it's a sanity check that the score behaves as designed, not a behavioral finding. “Does today's score predict the next 40 days” is the direction that isn't guaranteed and is where the real content is. We report both, and say so explicitly, so a mechanical relationship doesn't get mistaken for a discovered one.

Two forecast variants, one summarized here

We tested both an acceleration-only variant (the score's own rate of change, sensitive to turning points) and a blend of trend-level and acceleration (sensitive to both sustained trend and turning points), matching a toggle in the original systematic-trading construction. Both gave qualitatively similar results throughout; this article reports the blend variant only, for simplicity.

Windows, and what they don't fix

The 5/10/20/40-day windows are cumulative sums, not a day-by-day lag grid, chosen specifically because single-day flow is noisy enough that a real relationship can be invisible at daily granularity and only show up once averaged over a window - which is exactly the pattern we saw. What this doesn't fix: the four windows for the same product overlap heavily with each other (a 40-day window fully contains the 20-, 10-, and 5-day ones), so they are not independent draws. Treat the numbers here as descriptive tendencies across a broad set of products, not as formally significance-tested effects.

Data and coverage

Daily NAV, AUM, and flow (split-adjusted, via this project's existing flow-calculation pipeline), Tracking-basket only, over up to five years of history per product, trimmed to the roughly three years most products in this database actually have clean coverage for. Of 221 Pareto and 1,211 Leveraged/Inverse products in the corresponding dashboard universes, 189 and 604 respectively had enough history (at least the ~330 trading days needed to warm up the slowest moving average and volatility estimate) to include. Leveraged/inverse's much higher attrition rate reflects how many of these products are recent launches, particularly single-stock leveraged funds.

References

1. Carver, R., 2015, Systematic Trading: A unique new method for designing trading and investing systems, Harriman House.

2. Avellaneda, M., and Zhang, S., 2010, Path-dependence of leveraged ETF returns, SIAM Journal on Financial Mathematics 1(1), 586–603.

3. Jain, P. K., Mishra, S., Pagano, M. S., and Rodriguez, I., 2024, Of seesaws and swings: the market-wide impact of levered ETF rebalancing during stressful times, working paper, presented at the European Financial Management Association 2024 Annual Meeting.