Systematic Municipal Fixed Income · Strategy White Paper

KEYSTONE

Dislocation-Reversion in Individual Municipal Bonds — buying forced-seller mispricings on the regulatory trade tape
Asset class: US tax-exempt municipals (individual CUSIPs) Data: MSRB EMMA regulatory trade prints Backtest: – Status: research prototype, reproducible
For institutional discussion only. This document describes a research prototype and a backtest on historical public trade data. It is not investment advice, an offer, or a solicitation, and past simulated results do not guarantee future performance. All figures are net of the modeled customer bid–ask spread but gross of fees, financing, and taxes. Assumptions and known limitations are stated explicitly in §9.
1 · Executive summary 2 · The opportunity & microstructure 3 · Data & validation 4 · Methodology — the honesty harness 5 · The signal 6 · Evidence 7 · Performance vs MUB 7.5 · KEYSTONE-XL refinements 8 · Risk, regime & capacity 9 · Honest limitations 10 · Implementation & reproduction

1Executive summary

Individual municipal bonds cannot be day-traded profitably — the dealer bid–ask spread eats every short round trip. But that same dealer-market friction creates a durable, harvestable inefficiency for a patient systematic buyer: when a customer is forced to sell a bond well below its own recent level, a rules-based strategy that buys the dislocation and holds it to reversion earns a statistically significant excess return over random entry in the same bonds.

Update — KEYSTONE-XL (§7.5): the execution stack validated on corporates (limit entry · issuer cap · recovery exit) transfers verbatim to this strategy and is now the recommended operating point: +7.82% CAGR at −4.8% max drawdown over 2012–2026 (base: +4.35% / −9.1%), OOS 2023–26 CAGR +9.29% at a 98% win rate (2026-08 audit: issuer cap now enforced on real issuers — see XL_AUDIT.md). The sections below document the base strategy; §7.5 documents the upgrade, its equity curve, and real example trades.

The strategy — buy when a bond prints ≥3 points below its own trailing 60-day median, hold ~1 year, sell into a customer bid — was validated on a purpose-built harness designed to make self-deception hard: fills only at prices real customers actually received, a matched random-entry control that nets out carry and bond selection, a pre-registered parameter grid, and a locked out-of-sample test. It is a mean-reversion / liquidity- provision strategy and carries the corresponding risk: it underperforms in a sustained rate selloff (2022). We disclose that failure mode prominently rather than optimize it away.

What a manager is buying

A repeatable, capacity-aware source of +3.25% per-trade excess return over a matched control (p<0.001) in a $4-trillion, structurally inefficient asset class where the incumbent liquidity provider (the dealer) captures the spread — here re-routed to the patient investor. Signal, execution model, and full backtest are reproducible from public data.

2The opportunity & microstructure

The US municipal market is roughly $4 trillion across ~1 million CUSIPs and ~50,000 issuers. Unlike equities, there is no central limit order book: bonds trade over-the-counter through dealers, most CUSIPs trade only intermittently, and retail investors buy at the dealer's ask and sell at the dealer's bid. Every trade is reported to the MSRB and republished on EMMA, giving a complete regulatory tape of prices, yields, sizes and sides.

Why naive timing fails

The median same-day customer round-trip spread in our data is ~0.44 points (mean ~0.9). A two-week trade earns ~0.2% of carry — it cannot cover a ~0.9-point spread. Short-horizon reversion and momentum signals all lose and none beats a random-entry control (§6).

Why patience wins

The spread is a one-time cost; carry plus pull-to-par is a growing return. Past ~6 months the average customer round-trip turns positive; by one year it is materially positive. The dislocation signal then adds selection on top of that base.

Fig 1 — Mean customer round-trip return (buy ask → sell bid, coupon credited) by holding horizon, unconditional across all bond-days. The break-even is ~6 months; the strategy holds ~1 year.

3Data & validation

Source: MSRB EMMA, the regulatory trade tape for US municipal securities. Every dealer must report every muni trade; EMMA republishes each print with price, yield-to-worst, par size and side (customer-buy, customer-sell, inter-dealer). This is tick-level truth, not a matrix/model price.

Independent validation

Before any backtest, the download was validated four ways. The decisive check is cross-endpoint agreement: per-security price ranges and trade counts recomputed from the trade files match EMMA's independent summary endpoint exactly, so the data is not corrupted in transit.

CheckResultReading
Cross-endpoint price/count match100%two independent EMMA endpoints agree exactly
Δprice / Δyield correlation−0.98price & yield fields internally coherent
Customer-buy ≥ customer-sell (same day)96.9%side codes read correctly; spread is real
Median round-trip spread0.44 ptsthe friction the strategy must clear

4Methodology — the honesty harness

The rules below were fixed before results were produced. Their purpose is to make a false positive hard to manufacture.

5The signal

KEYSTONE entry rule

Enter when a bond's customer-buy price prints ≥3.0 points below its trailing 60-day median mid (a forced-seller dislocation) and the bond clears the liquidity gate. Fill at the next customer-buy print. Exit at the first customer-sell print after a 365-day minimum hold (hard stop 455 days). Coupon accrues throughout.

The signal is deliberately own-history and price-based. A yield-based version is confounded by market-wide rate moves and adds little once the control nets out bond identity (§6). A price dislocation relative to the bond's own recent level cleanly captures idiosyncratic forced selling — fund redemptions, tax-loss selling, estate liquidations — which mean-reverts.

6Evidence

6.1 Short-horizon timing does not work

Reversion, momentum and post-issuance strategies that round-trip in days–weeks all lose to the spread and none beats its matched control, in-sample or out.

Family (short hold)OOS nWinMean / tradeExcess vs ctrlp

6.2 Cheapness is monotone — the depth of the dislocation matters

Sorting one-year holds by entry discount is cleanly monotone: the deeper the bond prints below its own trend at entry, the higher the forward return. The excess over the matched control rises with the threshold and becomes highly significant at ≥3 points.

Fig 2 — Forward 1-yr return by entry price-discount decile (0 = cheapest).
Fig 3 — Excess vs control by entry threshold, full sample.
Entry thresholdTradesWinMean / tradeControlExcessp
Full sample 2013–2025. "Excess" = strategy − matched random-entry control in the same bonds.

6.3 In-sample → locked out-of-sample

The ≥3-point config was locked mechanically from in-sample and run once on data it had never seen. It strengthened out-of-sample (a regime effect we dissect in §8), remaining highly significant.

WindowTradesWinMean / tradeExcess vs ctrlp

6.4 Regime robustness — the most important table

Split by era, the edge is present and significant in four of five regimes, with magnitude that scales with dislocation opportunity (largest in the COVID crash/recovery). It fails in 2022, the one sustained rate selloff — see §8.

Fig 4 — Excess return vs matched control by era, ≥3-point config. Red = the rate-selloff regime where the strategy loses.
EraTradesWinMean / tradeExcess vs ctrlp

7Performance — KEYSTONE-XL vs base vs MUB

The recommended operating point is KEYSTONE-XL (§7.5): the base dislocation strategy plus the corporate-validated execution stack. Growth of $1 over the full window, all books modeled as equal-weight portfolios of open positions:

Fig 5 — Growth of $1: KEYSTONE-XL vs the base locked config vs MUB, linear-attribution convention (munis do not print daily; see §9). Note the 2022 rate-selloff plateau (§8) — XL flattens where the base and MUB fall.
Total returnCAGRMax drawdown*Trades

*Drawdown understated by linear intra-trade attribution — see §9. Total return and CAGR are realized from actual entry/exit prints plus coupon.

7.5KEYSTONE-XL — the transferred refinement stack

The corporate research program (GRANITE-XL) developed three execution rules on corporates and validated them bidirectionally. Applied verbatim to this muni strategy — limit entry (fill only if the customer-buy price ≤ latest prior mid + 0.25), one position per issuer, and a recovery exit (sell at the first customer-sell print ≥21 days in once the day's mid recovers to the entry-day trailing median; 455-day stop) — each step improves the book in both windows:

BookTradesWinMean/tradeExcess vs ctrlCAGRMaxDD*
base (locked config), IS 2012–221,70163%+2.56%+1.48%+3.66%−9.1%
KEYSTONE-XL, IS 2012–2232573%+5.27%+2.66%+7.84%−4.8%
base, OOS 2023–262,27491%+6.64%+2.10%+5.69%−0.0%
KEYSTONE-XL, OOS 2023–2633698%+6.85%+2.37%+9.29%0.0%
*Linear intra-trade attribution (munis do not print daily); understates crisis paths — see §9. All excess figures vs matched random-entry controls facing the same fill cap; all p<0.001. The 2023–26 OOS window was a benign muni tape (near-zero drawdown for all variants).

In-sample the stack roughly doubles CAGR (+3.66%→+7.98%) while halving drawdown; out-of-sample it adds +3.75pp of CAGR at a 98% win rate on a quarter of the trades. The same mechanism as corporates: the limit cap keeps only fills where the dislocation is still fully available, and the recovery exit harvests the reversion and recycles capital (holds ~380→234 days). Full validation trail: munis/research/results/keystone_xl.json and corps/NEW_STRATEGY_SEARCH.md.

Anatomy of a trade — historical examples

Real KEYSTONE-XL trades with the bond's full surrounding tape. Selected by percentile of realized return (90th / median / 10th) among tape-dense trades — representative outcomes, not cherry-picked winners. Gold dots: days customers bought; slate dots: days customers sold; dashed line: the trailing 60-day median defining the dislocation.

Quant appendix — KEYSTONE-XL

Everything under the hood: where each point of return comes from, the shape of the trade distribution, monthly moments, the worst episodes, and what the book actually holds. The decomposition is an exact identity — the four components sum to the per-trade total.

Return decomposition (mean per trade)

ComponentMeanShare

Carry = coupon accrual over the hold. Entry edge = discount captured vs the entry-day mid (the limit cap's contribution). Reversion = mid-to-mid price recovery — the alpha engine. Exit cost = selling at the bid below mid.

Per-trade return distribution

Monthly return moments

Worst drawdown episodes

PeakTroughRecoveredDepth

Book structure & execution

Year by year

YearBook returnTrades enteredMean/trade (entry vintage)

Execution & live replication — KEYSTONE-XL

Already inside every backtest number: fills pay the full printed bid–ask (buy at the customer-ask VWAP, sell at the customer-bid VWAP); no fill unless the required side actually printed (no phantom liquidity, 7-day fill window or no trade); entry is the day after the signal (no latency assumption); dead bonds exit at their last real bid. What is not automatic is competition for the print and size — modeled below.

Slippage stress — pay up h to win the print

Haircut hMean/tradeWinCAGRMaxDD

Capacity at 25% participation of printed volume

Position capped at 25% of the entry day's printed volume in that bond. Conservative: a patient buyer can also work orders across the multi-day dislocation window, and reported TRACE volumes are top-coded (large trades shown capped), so true capacity is higher.

Live-desk checklist — what a replication must handle

8Risk, regime & capacity

Known failure mode: sustained rate selloffs

KEYSTONE is short volatility / short a rate shock. In 2022, muni yields repriced up for a full year; "3 points below the 60-day median" then fired on every bond as the whole market fell, with no reversion — the strategy bought falling knives (−1.99% excess, 32% win). This is intrinsic to mean-reversion, not a fixable bug. Size it accordingly.

Regime overlay (optional)

Gating entries on the broad market (only buy when MUB ≥ its 100-day average) removes most of the 2022 loss and lifts full-sample excess to ~+4.3%, but cuts trade count ~6×. We present it as a risk overlay, not baked into the core — filtering on the very event that motivated it invites curve-fitting, so both versions are disclosed.

Capacity & positioning

Deep idiosyncratic dislocations are rare — on the order of ~230 per year in the sampled book, out of ~7,000 bonds trading ≥250×/yr nationally — and each fill means being the buyer into someone else's forced sale. Coverage scales with the tape: the full EMMA universe is ~500k CUSIPs, so the opportunity set is far larger than the sampled backtest. This is a capacity-limited, patient-capital strategy, complementary to a core muni book, not a high-turnover sleeve. As of the last data date it is dormant: no bond trades ≥3 points below its own trend, exactly as a liquidity-provision strategy should behave in calm markets.

9Honest limitations

10Implementation & reproduction

The entire pipeline — data acquisition, validation, panel construction, backtest, config lock, out-of-sample run, and this document's data — is reproducible from public sources:

python scripts/build_universe.py # discover the tradable universe (EMMA) python scripts/download_trades.py # full per-CUSIP trade histories python scripts/validate_data.py # structural + cross-endpoint validation python research/panel.py # daily per-bond panel python research/run_backtest.py is # in-sample grid python research/lock_configs.py # mechanical config lock python research/run_backtest.py oos # one-shot out-of-sample python research/current_picks.py # today's live screen

Deployment for a live book: run current_picks.py daily against the fresh EMMA tape; entries surface only when forced selling creates a ≥3-point dislocation in a liquidity-eligible bond. Position sizing, credit screening, and an optional MUB-trend overlay sit on top of the raw signal.

KEYSTONE — municipal dislocation-reversion. Research prototype on MSRB EMMA public trade data. Full methodology, code, and per-era detail in the project repository (munis/research/FINDINGS.md). Generated from docs/munis_data.json; every figure recomputes from the trade panel.

For institutional discussion only · not investment advice · simulated past performance is not indicative of future results.