This study moved here from the freight demo’s main page; no figure changed.
Freight brokerage demo
How far can a freight broker trust an old rate?
A broker often prices a load from a rate it saw weeks ago. This demo measures, on public weekly quotes for refrigerated produce lanes, how often an older rate still lands inside the current week’s published range. It is arithmetic on history: a fixed rule computes it, and no model is involved.
Demo, not validated · code only, no model
Data pulled
The result, on weeks held out until the run
Every figure comes from one pass over one frozen pull of public data. The method was explored on earlier weeks first; the headline figures are the weeks after, held out until the run. None of it is validated, and none of it is pricing advice.
A second demo, in which Jev reads carriers’ status messages and code decides what reaches a person, will be added to this page when it has run.
01 · Rate staleness
How often does an older rate still land inside this week’s range?
Demo, not validated · code only, no model
Brokers price from history: what a lane paid last week, last month, last season. Markets move, and fuel is often blamed. The US Department of Agriculture publishes a low and a high truck rate each week for refrigerated produce lanes into major cities. For each lane and week, the demo takes an older rate for the same lane and asks whether it lands inside this week’s range.
How it is measured
- A lane-week is one lane in one week. Where the report lists that lane more than once in a week, the demo takes the median of each published figure: the low, the high and the midpoint.
- The benchmark is the same lane’s midpoint an exact number of weeks earlier: 1, 4, 13, 26 and 52 weeks.
- The fuel-adjusted benchmark adds the change in the national weekly diesel price, divided by 6.5 miles per gallon, times the lane’s miles.
- The last 10 rates, averaged: the mean of the lane’s previous 10 weekly midpoints, however far apart they are.
- A benchmark is in range when it lies between the week’s low and high, inclusive. Its error is the gap to the midpoint, as a share of the midpoint.
- The method was explored on to . It was then run once on the weeks from to , which were held out until then: 11,610 lane-weeks on 213 lanes.
The same quote, carried forward
The published quote for a lane is often repeated unchanged from one week to the next. Where it was carried forward, an older rate lands inside the range almost by construction: every one of the 6,432 week-old rates carried forward did. So the short-horizon figures partly measure how often the quote is updated, not how the market moved. Each one- and four-week figure shows the carried-forward share beside it, and the in-range share over changed quotes only. Those columns were added after the predictions were registered.
The result, by the age of the rate
- Same quote carried forward, in range
- Changed quote, still in range
- Explored period, all quotes
Held out: 11,610 lane-weeks on 213 lanes. Each range is a lane-level bootstrap, from resampling whole lanes 2,000 times; it was added after registration.
| Age of the rate | In range, held out | Carried forward | In range, changed quotes only (added after registration) | Median error | In range, explored period |
|---|---|---|---|---|---|
| 1 week | 9,624 of 11,233 (85.7%, range 83.1 to 88.0%) | 57.3% | 3,192 of 4,801 (66.5%, range 62.7 to 70.1%) | 0.0% (range 0.0 to 0.0%) | 87.6% (range 86.5 to 88.6%) |
| 4 weeks | 7,909 of 10,689 (74.0%, range 70.5 to 77.4%) | 35.4% | 4,130 of 6,910 (59.8%, range 56.0 to 63.2%) | 2.7% (range 1.9 to 3.4%) | 68.8% (range 66.4 to 71.2%) |
| 13 weeks | 5,634 of 9,453 (59.6%, range 55.0 to 64.4%) | 21.0% | 3,647 of 7,466 (48.8%, range 45.2 to 53.1%) | 5.8% (range 4.7 to 6.9%) | 50.0% (range 47.0 to 53.1%) |
| 26 weeks | 4,449 of 8,442 (52.7%, range 47.2 to 58.7%) | 17.0% | 3,017 of 7,010 (43.0%, range 39.3 to 47.3%) | 8.0% (range 6.3 to 9.3%) | 40.4% (range 37.4 to 43.5%) |
| 52 weeks | 4,984 of 9,667 (51.6%, range 47.5 to 55.8%) | 12.0% | 3,827 of 8,510 (45.0%, range 41.6 to 48.6%) | 7.8% (range 6.8 to 8.8%) | 37.3% (range 35.3 to 39.4%) |
| Last 10 rates, averaged | 7,919 of 11,595 (68.3%, range 64.6 to 71.7%) | not applicable | not applicable | 3.4% (range 2.7 to 3.9%) | 59.0% (range 56.5 to 61.5%) |
The held-out weeks behaved differently
From thirteen weeks on, an older rate landed in range more often in the held-out weeks than in the explored period: at 52 weeks, 51.6% against 37.3%. The rule did not change, so this is not the method doing better on new data; the weeks themselves differed. The extract shows three differences: the published band is wider (13.3% of the midpoint against 11.1%), more thirteen-week-old quotes were carried forward (21.0% against 15.7%), and fewer lanes report each year (176 in 2024, 145 in 2025 and 102 in 2026 to , against 213 in 2023). Over changed quotes only, a thirteen-week-old rate still landed in range more often (48.8% against 40.6%), so none of these accounts for it alone. The reason is not established.
Does adjusting for diesel help?
| Age of the rate | Carried forward | Median error | Fuel-adjusted | In range | Fuel-adjusted in range |
|---|---|---|---|---|---|
| 1 week | 57.3% | 0.0% (range 0.0 to 0.0%) | 0.4% (range 0.3 to 0.6%) | 85.7% (range 83.1 to 88.0%) | 82.1% (range 78.7 to 85.0%) |
| 4 weeks | 35.4% | 2.7% (range 1.9 to 3.4%) | 2.9% (range 2.1 to 3.7%) | 74.0% (range 70.5 to 77.4%) | 68.8% (range 65.0 to 72.6%) |
| 13 weeks | 21.0% | 5.8% (range 4.7 to 6.9%) | 6.0% (range 4.9 to 7.1%) | 59.6% (range 55.0 to 64.4%) | 54.5% (range 49.8 to 59.3%) |
| 26 weeks | 17.0% | 8.0% (range 6.3 to 9.3%) | 7.4% (range 6.0 to 8.6%) | 52.7% (range 47.2 to 58.7%) | 48.9% (range 44.3 to 54.0%) |
| 52 weeks | 12.0% | 7.8% (range 6.8 to 8.8%) | 7.5% (range 6.6 to 8.3%) | 51.6% (range 47.5 to 55.8%) | 48.6% (range 44.5 to 52.7%) |
At thirteen weeks, adding the diesel change raises the median error on the point estimates, from 5.8% to 6.0%, and the two ranges largely overlap; it moves the in-range share from 59.6% to 54.5%. Diesel is one part of a truck’s cost; what else moves these rates is below.
The last rates, averaged
The mean of the lane’s last 10 weekly midpoints lands in range for 7,919 of 11,595 (68.3%, range 64.6 to 71.7%), with a median error of 3.4% (range 2.7 to 3.9%). A week-old rate’s median error is 0.0% (57.3% of them carried forward), and a thirteen-week-old rate’s is 5.8%. Those 10 quotes span a median of 70 days, and a tenth of them span 287 days or more.
What moves the rate (descriptive)
The same weekly report scores how hard trucks are to find on each lane. The table gives the median change in the rate over four weeks, by the change in that score over the same four weeks. In lane-weeks whose rate moved by 10% or more, the diesel change accounts for a median of 3.4% of the move (2,162 lane-weeks). This section is descriptive only: no prediction covered it, and its medians carry no range. Rows with fewer than 100 lane-weeks are marked thin; read them with care.
| Change in the availability score over four weeks (positive: trucks scarcer) | Lane-weeks | Median change in the rate |
|---|---|---|
| −3thin: few lane-weeks | 20 | −9.8% |
| −2 | 439 | −7.2% |
| −1 | 1,099 | −2.0% |
| 0 | 7,629 | 0.0% |
| +1 | 1,048 | +7.3% |
| +2 | 454 | +5.3% |
What was predicted, and how it came out
The lab’s research session registered four predictions after it had seen the explored period’s figures and before any held-out figure existed. They are quoted here as registered, so each is marked informed. The scores are the research session’s, not this page’s.
| Prediction | As registered | Confidence | Measured | Result | Against the lane-level range (added after registration) |
|---|---|---|---|---|---|
| 91 | A one-week-old rate lands inside the week's quote range in at least 85% of held-out lane-weeks.Informed: the research session explored 2001 to 2023 with the same method before registering. | 65% | 85.7% | Hit | hit on the point estimate; the lane-level range (83.1% to 88.0%) crosses the line. |
| 92 | A thirteen-week-old rate lands inside it in between 40% and 65% of held-out lane-weeks.Informed: the research session explored 2001 to 2023 with the same method before registering. | 65% | 59.6% | Hit | hit, marginal at the upper edge. The lane-level range (55.0% to 64.4%) stays inside the 40% to 65% band, but ends 0.6 points short of 65%. |
| 93 | At thirteen weeks, the fuel adjustment lowers the median error by less than one percentage point, or raises it.Informed: the research session explored 2001 to 2023 with the same method before registering. | 75% | 6.0% against 5.8% | Hit | hit, holds across the range. The extract carries each median error's range: 4.7% to 6.9% as quoted, and 4.9% to 7.1% fuel-adjusted. It doesn't carry the range of their difference. The research session's paired lane-cluster bootstrap of that difference, with the same settings, ran from −0.07 to +0.38 points, below one point in every resample. |
| 94 | The last ten rates have a lower median error than a thirteen-week-old rate, and a higher one than a one-week-old rate.Informed: the research session explored 2001 to 2023 with the same method before registering. | 75% | 3.4% against 0.0% and 5.8% | Hit | hit, holds across the range. The last ten's lane-level range (2.7% to 3.9%) doesn't overlap the thirteen-week one (4.7% to 6.9%), and the one-week median error is 0.0% throughout. |
The research session on 91: A hit, but mostly a measure of how often USDA carries a quote forward: most of these one-week comparisons repeat last week's quote unchanged (57.3%). Over changed quotes only, a one-week-old rate is in range less often (66.5%). That measure was added after registration.
The research session on 94: The lower side was near-certain, because most one-week quotes are carried forward unchanged.
For scale: across all the lab’s demos, the research session’s score stands at 53 hits of 76 predictions scored.
What this shows
- A fixed rule, arithmetic on public history, can say how stale a rate is, with no model and at no cost.
- Over the held-out weeks, about six in ten thirteen-week-old rates landed inside the week’s range (5,634 of 9,453). A week-old rate landed inside for 85.7%, but in 57.3% of those comparisons the quote had simply been carried forward; over changed quotes only, 66.5%.
What it doesn’t
- Refrigerated produce lanes into major cities only: not dry van or flatbed, and not the lanes any one broker runs.
- Published market quotes, not a broker’s buy rate or one carrier’s history.
- Thinning coverage: lanes reported 176 in 2024, 145 in 2025 and 102 in 2026 to .
- One pass, on one frozen pull of the data; the held-out weeks are fixed by that pull.
- In range means inside the week’s published band, a median 13.3% of the midpoint wide: near the market, not an exact price.
- Time and money saved: not measured. A saving needs the minutes a person spends pricing a load and what that time costs, and this demo measures neither.
02 · What a real deployment would need
From a public benchmark to a broker’s desk
- The broker’s own buy rates and lane history, not published market quotes.
- Dry van and flatbed lanes, and the lanes the broker actually runs.
- A record of when each quote last changed, so a carried-forward quote is not read as a fresh one.
- A test on weeks no setting was chosen on, as here, repeated as the market changes.
- A person setting the price on every load, with the rule as a check on staleness, not the price itself.
03 · How far to trust these numbers
One pass, one frozen pull
Every figure comes from one deterministic pass of a fixed rule over one pull of two public datasets. Every share carries a ninety-five per cent range from resampling whole lanes, not single lane-weeks: weeks on the same lane move together, so a range that treated each lane-week as independent would be too narrow and would overstate how sure these figures are. Those lane-level ranges were added after registration. The method and its constants were set on the explored period before the held-out weeks were run.
Where each figure comes from
Both datasets are public. The code that computed every figure is in the lab’s private repository, at the commit below, and is not public. This page is built from that run’s frozen extract, which holds counts and no lane names; the build fails if the extract’s counts disagree with each other. The extract itself is public: the frozen extract.
| Input | Dataset | Downloaded | Last week in the data | File |
|---|---|---|---|---|
| Refrigerated truck rates, US Department of Agriculture | acar-e3r8 | sha256 b8cbbf765189b99720b95697f6db5c6316adac6767c0d7d32eb33adf13ca304a | ||
| Weekly on-highway diesel price, US Energy Information Administration | EMD_EPD2D_PTE_NUS_DPG | sha256 7c5b0ac9885032af2bafe7827af7bece9662e97b5823b1c0b556af329ee96d48 |
The code that ran: 5a7661e27fa33f2ffc592a5f93953ba7748aa8ed
Cost: no model was called and nothing was paid. Both datasets are public.
Disclosure: this demo called no model. The lab has no relationship with TypeSafe, the company that makes Jev, the model the lab’s other demos test. The lab’s owner designed, ran and judged this demo, working with AI coding assistants; nothing here has been replicated independently.