Two Calendar Effects at the Month Boundary
Bond seasonality meets rebalancing flows — two edges sharing one calendar window.
This article examines two distinct effects that share the same calendar window and the same tickers. The first is a pure bond seasonality: TLT tends to weaken in the first week of each month and rally in the last few days, regardless of what equities do. The second is a conditioned reversal trade: when stocks outperform bonds during the first half of the month, the underperformer tends to recover at month-end — consistent with institutional rebalancing flows. Both operate on VTI and TLT, both use 5-day holding windows, and both are structurally different from the intraday sentiment patterns we explored in our Turnaround Tuesday and Friday Gold articles.
Two Effects, One Window
The bond seasonality — a standalone calendar pattern in which TLT rallies in the last days of each month — has been documented by practitioners including Kris Longmore at Robot Wealth, whose work was an important reference for this analysis. It requires no conditioning signal and no view on equities. The reversal trade is different: it rests on a rebalancing hypothesis. Pension funds, insurance companies, and balanced-mandate managers hold some variant of the classic 60/40 portfolio. When equities rally during the month, the portfolio drifts away from its target weights. At month-end, the mandate requires action: sell what outperformed, buy what underperformed. These flows are predictable in timing, directionally biased, and large enough in aggregate to create temporary dislocations in both equities and bonds.
We analyse both effects together because they share the same instruments, the same 5-day holding windows, and often overlap in calendar time. Separating them is essential for understanding which edges are unconditional and which require a conditioning signal.
As Longmore candidly described in a year-end retrospective, he stopped trading the bond seasonality during a 2023–2024 flat stretch — and missed a strong recovery in 2025. The lesson resonates with everything we have discussed in this series: a strategy with a Sharpe ratio near 1 can easily spend a year or two underwater without anything being broken. That is not a failure of the edge. It is exactly what variance looks like at this Sharpe level.
Anatomy of a Monthly Cycle
All analysis in this article uses actual ETF data from August 2002 onward — the first full month after TLT’s inception on July 22, 2002. VTI has been live since May 2001. No synthetic or backfilled data is used. The chart below traces the average cumulative return of both TLT and VTI from the previous month’s close through each successive trading day, averaged across 273 months through June 2025.
The pattern is remarkably clean. TLT drifts steadily downward through the first seven trading days, reaching a cumulative trough near −0.30%. It then flattens mid-month and rallies sharply in the final three to five days. VTI does the opposite: it captures the bulk of its monthly return in the first few days — the well-documented turn-of-month equity effect — and largely stalls thereafter.
The day-level view confirms the asymmetry. Counted forward from the start of the month, trading day 3 stands out as the weakest single day for TLT (t = −2.9), and the entire first week is consistently negative. The recovery begins around trading day 19 and accelerates into the final days.
Counting backward from month-end sharpens the picture, because months vary in length and the rebalancing effect anchors to the month boundary rather than a fixed calendar date.
How Long to Hold
Before introducing any conditioning signal, it is worth scanning the standalone calendar effect across different holding periods. The chart below shows the unconditional annualised Sharpe ratio for three simple strategies — long TLT over the last K days of the month, long VTI over the first K days, and short TLT over the first K days — for holding periods from 1 to 7 days.
The end-of-month long TLT effect is strong across all holding periods tested, peaking near 1.0 at 2–3 days and remaining above 0.8 through 7 days. The beginning-of-month short TLT is marginal unconditionally — its Sharpe only exceeds 0.4 beyond 5 days. The beginning-of-month long VTI sits in between. We adopt a uniform 5-day window for all sleeves — long enough to capture the effect without overfitting to a single trading day, and short enough to keep execution simple.
The Relative Performance Signal
The raw calendar profile tells us when the effect concentrates. The rebalancing hypothesis tells us more: the direction and magnitude of flows should depend on which asset outperformed during the month. If stocks led bonds in the first half of the month, rebalancers need to sell equities and buy bonds. The reverse if bonds led.
To test this directly, we measure the VTI−TLT return spread over the first 15 trading days of each month and plot it against the same spread in the subsequent trade windows. This treats the relative performance itself as the target variable, rather than individual asset returns.
The end-of-month window shows a clear negative slope: months where stocks outperformed bonds in the first 15 days see a reversal in relative performance during the last 5 days. The R² is modest — this is a monthly frequency effect and individual months are noisy — but the direction is unambiguous and the regression is significant. The beginning-of-month window shows a positive slope, consistent with momentum continuation in the early part of the next month.
What Conditioning Reveals
To disentangle the bond seasonality from the reversal trade, we define a binary signal: stocks outperformed (S>B) or bonds outperformed (B>S) over the first 15 trading days. Stocks lead in roughly 61% of months. We then examine all five possible trade sleeves — long TLT and long VTI at end-of-month, long VTI, short TLT, and long TLT at beginning-of-month — each broken down by the unconditional average and by the two conditioned subsets.
The chart identifies which sleeves to trade and under which condition. End-of-month long TLT is strong unconditionally — the signal modulates magnitude, not existence. End-of-month long VTI only works in B>S months: the mean-reversion leg, where bonds led and equities snap back at month-end. Beginning-of-month short TLT requires S>B conditioning — unconditionally its mean is near zero, but conditioned it becomes highly significant. Beginning-of-month long VTI shows a positive conditioned mean in S>B months, which makes it a natural candidate alongside the short TLT leg. And beginning-of-month long TLT is negative in S>B months and only marginally positive in B>S months — there is no trade there.
From this analysis, four candidates emerge: EOM long TLT (unconditional), EOM long VTI (B>S only), BOM long VTI (S>B only), and BOM short TLT (S>B only). The next step is to test whether each of these edges is statistically real.
Is This Real?
With only 273 signal months and individual sleeves containing between 107 and 273 trades, parametric t-tests alone are not fully convincing. We therefore run a permutation test for each of the four candidate sleeves: 10,000 random draws of equally many 5-day return windows from the same ticker, computing the compounded (geometric) mean each time. The question is where the actual sleeve’s geometric mean sits in that null distribution.
Three of the four candidates pass comfortably. The end-of-month long TLT and beginning-of-month short TLT sit at the 99.9th and 100th percentile respectively. The end-of-month long VTI (B>S) sits at the 99.9th percentile as well. But beginning-of-month long VTI (S>B) fails: at the 93.2nd percentile with a permutation p-value of 0.068, it does not clear the 95% threshold. Its parametric t-statistic looks respectable in isolation, but the permutation test asks the harder question — whether these specific 166 windows beat random draws from VTI’s native 5-day return pool, which already carries a positive drift. The answer is: not convincingly enough. We exclude it from the final ensemble.
What the Sleeves Look Like
Before sizing the overlay, it is worth comparing each sleeve’s return distribution against the native 5-day returns of the same ticker. The histogram overlays below show the traded windows in navy against the full set of overlapping 5-day returns in gray, together with the distributional moments.
The edge shows up consistently as a location shift — the sleeve distributions are displaced to the right relative to their native counterparts, with higher win rates across the board. The end-of-month long TLT sleeve shifts the mean by +0.33 percentage points per trade while slightly compressing the tails relative to native TLT: its kurtosis is 0.3 compared to 2.0 for all 5-day TLT windows. The end-of-month long VTI sleeve in B>S months shows the largest location shift (+0.68 pp) but also higher dispersion, with positive skew of +1.51 — the conditioned window selects for months where equities snap back. The beginning-of-month short TLT sleeve shifts the mean by +0.68 pp and exhibits near-zero excess kurtosis, consistent with a clean calendar effect without tail concentration.
The Overlay
We size each sleeve at 20% of portfolio capital. In months where stocks outperformed, the end-of-month window holds 20% long TLT and the beginning-of-month window holds 20% short TLT — never both at once. In months where bonds outperformed, the end-of-month window holds both 20% long TLT and 20% long VTI simultaneously, reaching a peak gross notional of 40%. We call this Version A.
Version B caps the overlap: in B>S months, the two end-of-month sleeves are reduced to 10% each, keeping the peak gross at 20%. This sacrifices some return for a tighter drawdown. Both versions spend 62% of all trading days entirely in cash.
These are overlay returns. The strategy adds 2.2% to 2.8% of annual return to whatever portfolio it sits on top of, at a volatility contribution under 2% and with drawdowns that have never exceeded 2.3% over 23 years of live ETF data. The Sharpe ratios are nearly identical between versions — reducing the B>S EOM overlap trades off 60 basis points of return for 60 basis points of tighter drawdown. The Calmar ratio slightly favours the reduced version. One note on implementation: we do not simulate borrowing costs for the short TLT sleeve. In practice, the short exposure can be implemented by temporarily reducing an existing long TLT position in the core portfolio rather than opening a separate short, which avoids borrowing entirely. We also do not model slippage, since all entries and exits occur at the close and can be executed as market-on-close orders.
Two Effects, One Overlay
The turn-of-month effect is structurally distinct from everything else in our seasonal toolkit. Turnaround Tuesday, the Friday gold reversal, and the FOMC drift are all sentiment-driven patterns that resolve within a session or two. The turn-of-month effect operates on a longer timescale and is driven by the mechanics of institutional rebalancing — trillions of dollars in mandated allocations that must be adjusted at predictable intervals. The signal does not predict anything. It identifies which direction the forced flows will push, and when.
The two effects examined here — the unconditional bond seasonality and the conditioned reversal trade — are distinct in origin but complementary in practice. The bond seasonality is a pure calendar anomaly that requires no signal and no view on equities. The reversal trade is structural, driven by the mechanics of institutional rebalancing, and only activates when the first-half relative performance provides a directional signal. Together they form an overlay that contributes 2–3% of annual return at under 2% volatility, adding a fundamentally different return driver to the seasonal toolkit we have been building throughout this series.
In the next article, we will step back from individual effects and assemble the full picture: how these seasonal overlays combine with the three-asset inverse-volatility core into a single, integrated portfolio.
https://stevenrob1.substack.com/