Trades & Field Services · Enquiry sizing
Large jobs reach an estimator first, not last
A sizing step that reads each enquiry and returns a likely price band and range, tested on 80,327 alteration jobs filed after the model was built.
Executive summary
Sorting enquiries by size read from the description puts 61% of total job value in the first fifth of the queue, against 42% for the typical cost of that trade and borough, and finds 62% of $250k+ jobs there.
Situation
Estimators work enquiries in arrival order, so large jobs can wait behind small ones and go to the firm that called back first.
Approach
A text model on the job description, building type and location, trained on 59,920 earlier jobs and tested on 80,327 later ones, with calibrated 80% ranges.
Result
Rank correlation with declared cost rises from 0.30 to 0.66. Ranges are well calibrated at 79% coverage but wide, so the output orders the queue and gives a first band rather than a quote.
The problem
A builder, fit-out firm or building-products supplier gets new work from two places: enquiries through the website and phone, and project leads from building approvals through services such as Cordell Connect. An estimator can price only a few a day properly. Firms that filter and rank leads before they reach the estimator say the gain is in reaching the large, suitable jobs first.
A precise quote from a paragraph of text is not realistic. What a job description does carry is a strong signal of scale, and I set out to measure how well that signal sorts jobs by likely size.
- Team
- Estimators, sales coordinator, operations manager
- Systems
- Website enquiry form, Cordell Connect or council approval data, HubSpot, ServiceM8 or simPRO
- Measure
- Share of total job value reached by the first enquiries worked; price band placed right
What I built
A sizing step that reads the enquiry description with the type of building and location, and returns a likely price band and an 80% range. Enquiries are sorted by expected size, so estimators reach the large jobs first and the customer can be given a realistic first band on the day.
Results
On 80,327 alteration jobs filed from July 2018 to December 2019, working the highest-sized 20% of enquiries first reaches 60.7% of all job value. Sorting by the typical cost for that trade and borough reaches 41.7%, and arrival order reaches 20%.
Price bands and ranges
| Description model | Trade and borough | |
|---|---|---|
| 80% ranges that contained the declared cost | 79.4% | 81.7% |
| Median range, top over bottom | 17.2× | 55.5× |
| Median error of the single figure | 59.5% | 71.9% |
| Single figure within 25% | 21.5% | 17.0% |
| Rank correlation with declared cost | 0.66 | 0.30 |
The 80% ranges hold their promise, containing the declared cost 79.4% of the time, but they are wide: the top is typically 17 times the bottom. That is a ranking and triage tool. Any figure given to a customer is a band, not a quote.
Where it goes wrong
| Declared cost fifth | Test jobs | Range | Model median error | Baseline median error |
|---|---|---|---|---|
| Lowest fifth | 16,223 | $1,000 to $12,500 | 157.5% | 438.2% |
| Second | 15,908 | $12,502 to $28,406 | 51.1% | 45.8% |
| Middle | 16,065 | $28,413 to $55,106 | 39.3% | 33.3% |
| Fourth | 16,448 | $55,110 to $150,000 | 49.1% | 56.3% |
| Highest fifth | 15,683 | $150,001 to $20,000,000 | 73.2% | 88.0% |
The smallest jobs carry the largest percentage errors, because a description such as plumbing work could be $2,000 or $20,000. Those enquiries matter least for estimator time, which is why the sizing is used to order the queue rather than to price.
Data and method
- Data
- NYC Department of Buildings job application filings (NYC Open Data, public): Alteration Type 2 jobs, which cover multiple trades with no change of use, filed 2017 to 2019. The target is the estimated job cost declared at filing. Applicant and owner names are not downloaded. 179,270 filings; 170,701 kept after removing those with no cost or description
- Split
- By filing date. 2017 filings train (59,920), January to June 2018 calibrate the ranges (30,454), July 2018 onward test (80,327)
- Model
- Ridge regression on log cost, with TF-IDF word and two-word features of the description, one-hot building type, borough and work type, and numeric building details
- Ranges
- Split-conformal 80% ranges from the calibration filings, set separately by predicted price band
- Baseline
- Quantiles of declared cost for the same type of work and borough in training
- Description length
- median 20 words
Limitations
- Declared cost at filing is the applicant’s estimate, not the final contract price, and some applicants under-declare.
- New York building work is not Australian work. Rates, building stock and description habits differ, so the model is retrained on a firm’s own enquiries and won or quoted values.
- A trade business’s enquiries are shorter and less technical than permit descriptions; accuracy is measured on its own history before use.
- The ranges are wide by design. They support ordering and a first conversation, never a binding price.
How I would take it live
- Export two years of enquiries with the value of the quote or the invoiced job
- Retrain with the firm’s own job types and suburbs and agree the bands with the estimators
- Show the band beside each new enquiry for a month without changing the order of work
- Then sort the estimator queue by expected size and add a same-day first-band reply