FOMC · reaction function

Ordered logit on 36 years of decisions · point-in-time data · walk-forward scored

10/19/2026, 11:39:33 AM

Frozen snapshot. The engine does not run on this host. These are the numbers cli.py produced on 10/18/2026, 5:15:31 PM, not a live read.

Incoming data since then is not in this page, and the futures read below is that day's pricing, not today's.

calibrated to

2.2pp

mean gap, said vs happened · 233 meetings

vs. fed funds futures, on moves

−17.3pp

65.8% called, 6.9% false alarms · 83.1% priced, 5.3% false alarms · n=65 moves

walk-forward window

233meetings

1997-08-19 → 2026-09-16

Next meeting 2026-10-28

10 days away

headline · fed funds futures

hold 84.0%

14 days out, the futures strip called direction on 88.4% of 164 meetings against the model's 86.6% (3 more right).

The two disagree: the model says hike, the market prices hold.

model

cut50+0.0%

cut250.0%

hold18.3%

hike2579.3%

hike50+2.4%

P(cut) 0.0%P(hold) 18.3%P(hike) 81.7%

current target (upper)4.00%

expected move+21.3 bp

implied target4.21%

Read the probability, not the argmax. The expected move is the distribution's mean in basis points, not a prediction that the Fed will move by that amount — the Committee only chooses in 25bp steps.

fed funds futures · headline

cut0.0%

hold84.0%

hike16.0%

priced move+4.0 bp

call (±12.5bp band)hold

CME ZQ strip through the FedWatch arithmetic, the same number the scorecard grades. Options (Atlanta Fed) put P(hike) at 93.0% and P(cut) at 2.1%, but over the nearest option window, not this meeting.

Direction called, walk-forward, by days before the meeting opens, on the meetings the strip reached that day. Recall on moves and false alarms on holds read together.
days out	n	model	market	moves m / mkt	false alarm m / mkt
1	196	84.7%	90.8%	64.6% / 83.1%	5.3% / 5.3%
7	187	84.0%	95.7%	62.3% / 95.1%	5.6% / 4.0%
14 ← now	164	86.6%	88.4%	62.0% / 86.0%	2.6% / 10.5%
21	132	84.8%	87.1%	57.9% / 81.6%	4.3% / 10.6%
28	130	85.4%	86.2%	55.6% / 91.7%	3.2% / 16.0%
42	67	83.6%	83.6%	50.0% / 77.8%	4.1% / 14.3%

CME watch — the priced path

Fed funds futures, meeting by meeting: each row reads its own ZQ contract through the FedWatch arithmetic, splits the priced move across the two 25bp outcomes that bracket it, and carries every earlier path forward. Unlike the conditional path below, this is a rate path: the one the market has priced (as of the snapshot date, not today). Read it as risk-neutral pricing, not a forecast: the two-outcome split assumes a shape the strip does not carry, so tails are understated. The options check under the table shows where that bites.

The move priced at each upcoming meeting, the contract it was read from, and the cumulative probability the target range is above today's.
meeting	contract	priced here	cut	hold	hike	implied EFFR	most likely	above 375-400
2026-10-28	ZQX2696.080	+4.0bp	0.0%	84.0%	16.0%	3.920%	375-400	16.0%
2026-12-09SEP	ZQZ2695.920	+22.5bp	0.0%	9.8%	90.2%	4.145%	400-425	91.8%
2027-01-27	ZQG2795.765	+9.0bp	0.0%	64.2%	35.8%	4.235%	400-425	94.7%
2027-03-17SEP	ZQH2795.690	+16.6bp	0.0%	33.6%	66.4%	4.401%	425-450	98.2%
2027-04-28	ZQK2795.515	+8.4bp	0.0%	66.4%	33.6%	4.485%	425-450	98.8%
2027-06-09SEP	ZQM2795.455	+8.6bp	0.0%	65.7%	34.3%	4.571%	450-475	99.2%
2027-07-28	ZQQ2795.405	+2.4bp	0.0%	90.3%	9.7%	4.595%	450-475	99.3%
2027-09-15SEP	ZQU2795.390	+3.0bp	0.0%	88.0%	12.0%	4.625%	450-475	99.4%

Probability of each target range after each meeting, in percent, chained across meetings.
after	550-575	525-550	500-525	475-500	450-475	425-450	400-425	375-400
2026-10-28							16.0	84.0
2026-12-09						14.4	77.3	8.2
2027-01-27					5.2	37.0	52.6	5.3
2027-03-17				3.4	26.3	47.3	21.2	1.8
2027-04-28			1.2	11.1	33.4	38.6	14.7	1.2
2027-06-09		0.4	4.6	18.7	35.1	30.4	10.0	0.8
2027-07-28		0.8	5.9	20.3	34.7	28.4	9.1	0.7
2027-09-15	0.1	1.4	7.7	22.0	33.9	26.1	8.1	0.6

Current range 375-400 (highlighted). EFFR going in 3.880%, the 10-day median. Probabilities under 0.05% are left blank.

cross-check · SOFR options (Atlanta Fed, report 2026-10-15)

A different instrument pricing the same question. Where the two agree, the strip read is corroborated. Where they part it is usually the tails: options put weight on a cut inside a hiking path, and a two-outcome split of the futures mean cannot.

Probability the target range sits above or below 375-400 by each option window, futures tree beside SOFR options.
by	after meeting	above · futures	above · options	below · futures	below · options
2026-12-16	2026-12-09	91.8%	93.9%	0.0%	2.4%
2027-03-17	2027-01-27	94.7%	88.0%	0.0%	6.1%
2027-06-16	2027-06-09	99.2%	84.5%	0.0%	10.1%
2027-09-15	2027-07-28	99.3%	79.5%	0.0%	14.6%

Federal Reserve Bank of Atlanta; CME Group market data used with permission. Personal and educational use only.

CME watch — the priced path

reading the fed funds futures strip across the next eight meetings…

walk-forward record

233 meetings, 1997-08-19 → 2026-09-16. Refit at every step on prior meetings only — scaler, imputation medians, cutpoints and coefficients, so no fit ever saw its own test row. Headline accuracy is meaningless on its own: the Fed held 68.7% of these meetings, so that is the floor, not the bar. Every figure below is read straight out of the engine's scorecard — this page does not recompute any of it.

On the meetings that matter, the engine loses to what was already priced.

Across the 65 meetings where the Fed moved and the futures strip reached (of 73 in all) , the fed funds futures called direction 83.1% of the time against this model's 65.8% — a gap of −17.3pp. And unlike the yield-spread proxy it does not buy that recall with false alarms: 5.3% of holds against the model's 6.9%. It is better on both axes at once. Treat the distribution as a structured read on why pressure sits where it does, not as an edge over the market.

Direction accuracy for the model and each baseline, over all scored meetings, over the subset where the Fed moved, and the false-alarm rate on the meetings where it held.
direction called by	scored on	all meetings	when it moved	false alarms on holds
this model	233	84.5%	65.8%	6.9%
always say hold	233	68.7%	0.0%	0.0%
repeat the last action	233	79.0%	65.8%	15.0%
market-implied (1y - funds proxy)	233	66.1%	91.8%best	45.6%
market-implied (fed funds futures)	196	90.8%best	83.1%	5.3%
model edge vs. best baseline		−6.3pp	−26.0pp	read across, not down

The "when it moved" column is recall and the column beside it is what that recall cost. A caller can buy the first with the second, which is exactly what the 1y-minus-funds proxy does. Coverage differs by row: the fed funds futures strip cannot price every meeting, and those meetings are left out of its denominator rather than scored against the Treasury proxy under its name. On the 196 meetings it does reach, every caller scores this model: 64.6% / 5.3% · always say hold: 0.0% / 0.0% · repeat the last action: 66.2% / 11.5% · market-implied (1y - funds proxy): 90.8% / 45.8% · market-implied (fed funds futures): 83.1% / 5.3% (moves / false alarms).

exact size, all meetings

82.4%

vs 68.7% for always-hold

exact size, on moves

58.9%

vs 0.0% for always-hold · n=73

false alarms on holds

6.9%

called a move when it held · n=160

mean |error|

6.8 bp

log loss 0.494 · Brier 0.263

calibration — is the distribution trustworthy?

Read this before any single forecast. Each point is a bucket of meetings the model assigned a similar P(move); the diagonal is perfect calibration. A point above the line means the Fed moved more often than the model said, below means it moved less. Scored walk-forward over 233 meetings.

calibration error (n-weighted)

2.2pp

mean gap between what it said and what happened

worst bucket

8.8pp

in the 40-60% bucket — an average can hide one broken bin

Calibration of the predicted probability of a policy move, by bucket.
model said	n	predicted	happened	gap
0-20%	141	7.3%	8.5%	+1.2pp
20-40%	21	28.3%	28.6%	+0.3pp
40-60%	20	51.2%	60.0%	+8.8pp
60-80%	12	71.0%	66.7%	−4.3pp
80-100%	39	92.9%	89.7%	−3.1pp

Buckets are the engine's own, from models/backtest.json. A bucket holding a handful of meetings is noise — read n before reading its gap.

conditional path — pressure already in the data

This is not a rate-path forecast. Do not quote it as one.

Every meeting past the first is scored on today's data. The model has no mechanism to forecast where CPI or unemployment will be in six months, so these rows answer “what would the Fed do at that meeting if nothing changed” — a read on how much tightening or easing pressure is already in the data, not a call on the terminal rate. The only row that is a forecast is the first one.

The reaction function evaluated at the next 5 scheduled meetings, all on today's data. Not a rate path.
meeting	basis	cut	hold	hike	exp. move	implied
2026-10-28	the forecast	0.0%	18.3%	81.7%	+21.3bp	4.21%
2026-12-09SEP	today's data	0.0%	4.3%	95.7%	+28.0bp	4.28%
2027-01-27	today's data	0.0%	2.1%	97.9%	+32.1bp	4.32%
2027-03-17SEP	today's data	0.0%	0.1%	99.9%	+57.9bp	4.58%
2027-04-28	today's data	0.0%	0.8%	99.2%	+39.9bp	4.40%

Implied targets are each meeting's expected move applied to the current target (4.00%) — they do not compound down the column, because each row is an independent read on today's data rather than a step in a sequence. Reading the last row as a terminal rate is the specific misuse this table is built to prevent.

conditional path — pressure already in the data

This is not a rate-path forecast. Do not quote it as one.

Every meeting past the first is scored on today's data — a read on how much pressure is already in the data, not a call on the terminal rate.

running the reaction function across the next five meetings…

what is driving this meeting

Exact arithmetic, not an approximation: each contribution is β_j · z_j, the fitted coefficient times the standardised feature value, in latent policy-pressure units. They sum to the latent score the cutpoints are applied to, so this decomposition IS the forecast rather than a story told about it.

pushing toward tightening

feature	value	β·z
mkt_3m	0.36	+1.17
ff_p_cut	0.02	+0.85
mkt_1y	0.59	+0.74
mortgage_chg_6m	0.86	+0.73
last_dir	1.00	+0.73
breakeven_5y	2.35	+0.64

pushing toward easing

feature	value	β·z
sentiment	51.70	−0.46
vacancy_ratio	1.00	−0.35
ff_p_hike	0.93	−0.33
months_supply	8.50	−0.30
months_since_move	1.28	−0.19
hpi_yoy	1.87	−0.12

the fitted reaction function

One signed coefficient per feature, biggest first — the thing a desk would actually argue with. The latent is policy pressure, so a positive β means a higher value of that feature pushes toward tightening. These are L2-penalised and fitted on standardised columns, so they are comparable to each other in size but are not elasticities, and a feature correlated with three others will have its weight split across them.

ff_exp_move_bp+1.412
mkt_3m+0.794
breakeven_5y+0.789
mkt_1y+0.765
chg_12m+0.712
mkt_2y+0.674
last_dir+0.565
credit_chg−0.520
mortgage_chg_6m+0.474
newsales_yoy−0.462
vacancy_ratio−0.411
ff_p_cut−0.368
nfci+0.339
months_since_move+0.330
slope_10y2y−0.321
claims_yoy+0.301
hpi_yoy+0.282
level−0.249

← pushes toward easingpushes toward tightening →

recent decisions

The most recent policy decisions, derived from the FRED target series.
2026-09-16	+25bp	hike25	3.75% → 4.00%
2026-07-29	0bp	hold	3.75% → 3.75%
2026-06-17	0bp	hold	3.75% → 3.75%
2026-04-29	0bp	hold	3.75% → 3.75%
2026-03-18	0bp	hold	3.75% → 3.75%
2026-01-28	0bp	hold	3.75% → 3.75%
2025-12-10	-25bp	cut25	4.00% → 3.75%
2025-10-29	-25bp	cut25	4.25% → 4.00%
2025-09-17	-25bp	cut25	4.50% → 4.25%
2025-07-30	0bp	hold	4.50% → 4.50%
2025-06-18	0bp	hold	4.50% → 4.50%
2025-05-07	0bp	hold	4.50% → 4.50%
2025-03-19	0bp	hold	4.50% → 4.50%
2025-01-29	0bp	hold	4.50% → 4.50%