Savers Co-Operative Bank (Southbridge, MA; RSSD 1002373) reported $766m in total assets at Q3 2026. The deposit betas below are estimated from 103 quarters of its Call Report filings, Q1 2001 to Q3 2026, and ranked against the 1,424 banks in its asset-size peer group ($300M – $1B).
Long-run deposit betas against effective fed funds, 10-year window (Q4 2016 to Q3 2026). Error-correction model, Newey-West standard errors.
| Product | Long-run beta | 95% CI | Rising | Falling | Peer median | Rank in peer group | Rate paid, Q3 2026 | Balance, Q3 2026 |
|---|---|---|---|---|---|---|---|---|
| Total interest-bearing deposits | 0.39 | 0.33 to 0.44 | 0.37 | 0.36 | 0.54 | #205 of 1,405 | 1.77% | $542m |
| Interest-bearing checking | -0.00 | -0.01 to 0.01 | -0.00 | -0.01 | 0.23 | #51 of 1,287 | 0.08% | $136m |
| Savings (incl. MMDA) | 0.26 | 0.16 to 0.35 | 0.21 | 0.14 | 0.39 | #396 of 1,311 | 1.21% | $199m |
| CDs ≤ $250k | 0.73 | 0.59 to 0.88 | 0.71 | 0.63 | 0.73 | #763 of 1,408 | 3.32% | $124m |
| CDs > $250k | 0.57 | 0.39 to 0.75 | 0.60 | 0.77 | 0.70 | #318 of 1,405 | 3.61% | $83m |
Rising and falling are the NARDL long-run betas on fed funds increases and decreases. Greyed figures have |t| below 2 and are not distinguishable from zero. Peer group: $300M – $1B in total assets, 1,424 banks; rank 1 is the lowest beta.
On total interest-bearing deposits, Savers Co-Operative Bank’s long-run deposit beta over the 10-year window to Q3 2026 is 0.39: a 100 basis point change in fed funds that persists settles into a change of about 39 basis points in the rate paid (95% confidence interval 0.33 to 0.44). That is below the peer-group median of 0.54 (#205 of 1,405, ranked from the lowest beta).
Rising and falling betas differ significantly at the 5% level on savings (incl. MMDA), so a single symmetric beta would misstate repricing in one direction.
Every window, estimator and fitted path is in the interactive analysis below. How these figures are estimated. Call Report figures are the outside view; for betas and decay rates from your own account-level data, email contact@depositanalysis.com.
Deposit rate. Quarterly interest expense ÷ quarterly-average balance × 4, the UBPR’s own convention for annualising a quarter. Schedule RI expense is reported year-to-date and is therefore differenced for Q2–Q4 and taken as filed in Q1. Schedule RC-K balances are already single-quarter averages of daily balances. Because quarters run 90 to 92 days, ×4 carries a few basis points of calendar variation at high rates (about ±5bp at 5%); it is common to every institution and every peer band, and small against the moves a beta is estimated on.
Total interest-bearing deposits includes foreign-office deposits, as the UBPR’s total interest-bearing deposit cost does — on the institution’s card and in its peer ranking alike.
The call report split time deposits at $100,000 through 2016Q4 and at $250,000 from 2017Q1. The reported items change with it: Schedule RC-K averages move from RCONA529 (under $100,000) and RCONA514 ($100,000 or more) to RCONHK16 ($250,000 or less) and RCONHK17 (more than $250,000), with Schedule RI interest moving from RIADA518/A517 to RIADHK03/HK04.
The two are chained into one series per bucket. Estimated on the $250,000 items alone a CD beta covers nine years and a single tightening cycle, which is not enough to distinguish a 10-year lookback from a 20-year one; chained, CDs are measured over 2001–2026 on the same footing as checking and savings.
The chain is continuous in sensitivity and discontinuous in level. On 2017Q1 the $100,000–$250,000 tranche moves out of the large bucket and into the small one: $351bn in aggregate, or 37% of the former $100,000-or-more bucket. At the median institution the effect is far larger — only about a third of that bucket remains in it — since the aggregate is carried by large banks whose time deposits are mostly over $250,000 in any case. None of this is repricing or runoff, and it is treated accordingly:
The sum of the two buckets is unaffected throughout, so the total and total-including-NIB series require no adjustment for the change. Where a CD balance series must be comparable in level rather than in sensitivity, the call report reports one directly: RCONJ474 (more than $250,000) and RCONJ473+J474 ($100,000 or more) keep a single definition either side of 2017Q1, both from 2010Q1 onward.
One question — the share of a fed funds move that reaches this deposit rate — estimated three ways. The estimates differ, and the difference is informative.
Regression measures speed: the movement in the quarter the change occurs. ECM · NARDL measures destination: the level at which the rate settles if the change persists. Cycle high-low is arithmetic rather than a model. An institution that reprices slowly but ultimately fully will show a low regression beta and a high ECM beta; both are correct, and ALM applications require the destination.
Neutral: rate = a + β·ff in levels — the deposit rate associated with a given fed funds level. This is the specification most commonly reported in published work and is included for comparability.
Rising / falling: first differences with a directional split, Δrate = a + βup(Δff×up) + βdown(Δff×down) + βlagΔfft−1, reported as βup+βlag and βdown+βlag. These are coefficients on changes — pass-through speed with one quarter of lag, not a settled level.
Limitations. The level regression places two trending series on either side of the equation, the standard spurious-regression setup. The difference betas are imprecise on low-beta products; on interest-bearing checking the rising beta frequently returns |t| below 2.
The error-correction model carries the level relationship and the quarterly dynamics in a single equation, rate = a + ρ·ratet−1 + β·ff + φ·fft−1, with long-run beta (β+φ)/(1−ρ). The lagged rate term permits short-run stickiness while identifying the settled level. This is the standard correction for the spurious-regression problem above and the specification used by the Federal Reserve Board in its own funding-beta research (Dias & Scott, FEDS 2025-083). Where ρ reaches 0.97 the deposit rate has no reliable equilibrium with fed funds — the divisor 1−ρ is close to zero, and at ρ ≥ 1 the long-run beta changes sign — so the beta is marked not usable, as are long-run betas outside −1 to +2. The error-correction t-statistic, (ρ−1)/se(ρ), is reported in the regression detail and the beta export: it tests whether a long-run relationship exists at all.
NARDL (Nonlinear Autoregressive Distributed Lag; Shin, Yu & Greenwood-Nimmo, 2014) is the asymmetric form, permitting the long-run beta to differ between increases and decreases. The fed funds path is decomposed into partial sums of its own increases and decreases:
ff⁺t = Σs≤t max(Δffs, 0) the hiking pathff⁻t = Σs≤t min(Δffs, 0) the cutting path
Both are levels, both carry a long-run relationship, and together they reconstitute the fed funds path. In error-correction form, rate = a + ρ·ratet−1 + β⁺ff⁺ + φ⁺ff⁺t−1 + β⁻ff⁻ + φ⁻ff⁻t−1, with long-run betas (β⁺+φ⁺)/(1−ρ) on increases and (β⁻+φ⁻)/(1−ρ) on decreases. Selecting a rising or falling cell returns a t-test of the difference between them.
Rationale for the asymmetric form. A symmetric ECM is weighted toward whichever regime dominates the sample. The symmetric beta falls between the rising and falling betas in only about 30% of fits, since 2001–2026 contains materially more easing and on-hold quarters than hiking quarters. Decomposing the path removes that bias. In a 500-bank random sample of current filers, re-estimated with the corrected standard errors, 33–51% of institution-product fits show asymmetry significant at the 5% level, depending on the window (both arms usable). The direction is not fixed. Over the full 2001–2026 history the falling beta is the larger in 85% of significant fits — deposit rates give back more of a cut than they passed through of a hike — while over the last ten years, which the 2022–23 hikes dominate, the rising beta is the larger in 86%. By product, over 20 years: savings and interest-bearing checking mostly fall further than they rise (82% and 67% of significant fits), while CDs split close to evenly and lean the other way. Which direction dominates therefore depends on the product and the window, and is read per institution rather than assumed.
ECM · NARDL is appropriate for ALM inputs. An ALM model applies a rate shock and holds it, so the relevant quantity is the level at which the deposit rate settles — the ECM long-run beta. The regression family reads low against this by construction: a one-quarter pass-through speed is not a full-shock pass-through and should not be substituted for one.
Regression is appropriate for comparability with published studies and as a cross-check. Where the two families diverge materially on a single product, the cause should be understood before either figure is used.
(Δ deposit rate) ÷ (Δ fed funds) between the endpoints of the most recent completed fed funds cycle. Both endpoints of both series are displayed, so the figure can be verified manually.
The endpoints are the fed funds series’ own confirmed turning points — a reversal of at least 25bp against the running high or low, so a single-quarter pause does not begin a new leg. The same two dates apply to every institution and product, which prevents arbitrary endpoints from landing on a near-zero fed funds change and inflating the ratio.
The measure rests on two observations: no t-statistic, R², confidence interval or window. It is a reference figure rather than an estimate.
Not identified. A window over which fed funds moved less than 100bp in total is not estimated: any slope fitted to a few basis points of noise is not a beta, however significant its t-statistic looks. The table shows n/i for these windows. Rising and falling quarters are those in which fed funds moved at least 5bp; quarters with both averages below 25bp sit at the zero lower bound and count as neither.
Standard errors. Newey-West (HAC, 4 lags) throughout, with the small-sample correction n/(n−k): without it a five-year NARDL’s standard errors are about a fifth too small. Combined rising and falling betas in the regression family are obtained from a joint linear-restriction test. Each ECM long-run beta is a nonlinear combination of three or more coefficients, so its standard error is derived by the delta method on the HAC covariance matrix — including the rising-minus-falling t-statistic, computed on the difference itself.
R². Reported per model, not per coefficient. Regression family: the level model (Neutral) and the dynamic model (shared by Rising and Falling). ECM family: the symmetric ECM (Base) and the NARDL (shared by Rising and Falling).
Only the level model’s R² describes the level of the rate; the remaining three describe the quarterly change, a materially harder target. An R² of 0.4 against a level R² of 0.95 is therefore expected, and the two are not comparable. The ECM and NARDL are estimated on the level with the lagged rate on the right-hand side; reported in that form they return 0.97–0.99 for nearly every institution, but an AR(1) containing no fed funds term returns approximately 0.85, and above 0.95 at large institutions. That statistic measures the persistence of deposit rates rather than the explanatory power of the rate path. Both are consequently reported in error-correction form, Δrate = a + (ρ−1)ratet−1 + … — identical equation, coefficients and residuals, scored against the change. Typical values are near 0.5.
Chart lines. Solid lines are reported data — deposit rates and effective fed funds. Dashed lines are fitted values (short: level; long: dynamic; tight: ECM), drawn only for models currently displayed and only when a single deposit type is selected. The cycle measure has no fitted path.
Noninterest-bearing. Pays no interest and is not rate-sensitive. It is displayed so that components reconcile to total deposits and is included in the “incl. NIB” cost series, which mechanically reduces any beta measured against that series.
Windows. Preset windows are nested and share a common end date at the latest quarter. Differences across them indicate regime sensitivity rather than independent estimates. A custom range is re-estimated server-side over precisely the quarters selected.
Repricing versus mix. Each institution’s card splits the change in its cost of interest-bearing deposits over the selected window into repricing (each product’s rate change at the starting mix), mix shift (the change in product weights at the starting rates) and their interaction, which sum exactly to the total. Mix shift is largely outside an institution’s control and does not reverse when rates fall; the headline beta cannot separate the two.
Peer group. Institutions are assigned to an asset-size band approximating the FFIEC UBPR cohorts, using total assets at the latest quarter, held constant across history. A single cohort for the full series is preferred to a moving assignment, which would compare a long history against a group the institution belonged to for only part of it.
Bands. Peer distributions are computed across the full cross-section of the cohort in every quarter and are published only where at least eight institutions report. A beta is ranked only where the institution’s own card would call it usable, from the same series and the same estimator, so the ranked figure and the card agree. Positions are expressed as ranks rather than percentiles: in cohorts of this size a single institution can be worth several percentile points, and a rank does not imply precision the sample cannot support.
Checking and savings are reported separately, following the Call Report itself: Schedule RC-K collects interest-bearing transaction accounts and savings deposits (including MMDAs) as distinct items, and both are actively reported — 4,148 and 4,242 of 4,336 filers respectively in the most recent quarter. They price differently, and an ALM model carries separate assumptions for each: across the filing population the median long-run beta is 0.17 on checking against 0.34 on savings.
One caveat follows from showing them apart. The April 2020 amendment to Regulation D removed the six-transfer limit on savings accounts, and institutions reclassified balances into interest-bearing transaction accounts in different quarters. About one filer in five moved more than five points of its deposit base between the two categories in a single quarter during 2020–2021. A step between these two bands in those years may therefore be a reporting reclassification rather than a change in funding. Their combined total is unaffected, so the remainder of the mix and every comparison against it are unchanged.
Reported. Schedules RC-B M.2 (securities), RC-C M.2 (loans), RC-E M.3 (time deposits) and RC-M 5 (borrowings) assign every dollar to a “remaining maturity or next repricing date, whichever is earlier” bucket. Floating-rate loans therefore fall in ≤3 months regardless of term. Securities comprise HTM at amortized cost plus AFS at fair value. Balances not covered by the ladders are held at zero duration, and the coverage share is disclosed.
Assumed. The Call Report reports balances, not durations. Each bucket is assigned one: midpoints converted to modified duration for coupon instruments, prepayment-adjusted for pass-throughs and residential loans — a 30-year pass-through carries 6.5 years rather than 30. The open-ended “over 15 years” buckets are the least precise, placing a 20-year municipal alongside a 30-year Treasury. The full table is available in the API response under meta.assumptions.
Deposit duration is solved, not assumed. Average life is not estimable from Call Report aggregates: net balance changes mask attrition and no vintage detail is reported. Duration of equity is therefore set to zero and the deposit duration that balances the sheet is derived: [D(assets)×assets − contractual liability duration$] ÷ NMD. Average life follows from the institution’s own ECM beta, life = duration ÷ (1 − β); a deposit repricing one-for-one carries no economic-value exposure regardless of how long the balance remains.
OCC medians determine shape, not level. The Interest Rate Risk Statistics Report publishes NMD average lives by asset-size band. Only the ordering across deposit types is used — noninterest-bearing longer than money market, and so on — scaled by a single factor per institution so that the weighted duration reconciles to the solve. The absolute level is calibrated on institutions below approximately $50bn and does not extend beyond that population.
Hedges. Schedule RC-L reports gross notional of non-trading interest rate swaps (RCONA589) and nothing further — no pay/receive split, tenor or hedged item. Pay-fixed at 4.6 years is assumed, the average bank swap term transacted in 2025. Above half of total assets the position is treated as an intermediation book rather than an ALM hedge and no directional assumption is applied. This is the least well-identified input on the page and the most readily resolved from an institution’s own confirmations.
Limitations. No account-level decay, hedge direction, coupon or vintage detail within a bucket, embedded caps and floors, or any information below the consolidated aggregate. This is a screening and benchmarking tool, not an ALM model, and is not a sufficient basis for a risk-limit decision.
Generated · Source: FFIEC Call Report and the Federal Reserve effective federal funds rate via FRED. For research use; not investment advice.