Internal scoring + opportunity engine. Two-person access. Authenticate via Google to continue.
Out of 280K owner-occupied CT homes you haven't served, these 5K most resemble your existing customer profile. T1 (top 1K) gets Concierge call, T2 gets direct mail, T3 gets digital. Each row has reason codes + estimated ticket.
Download CSV ↓Customers (red/gold, by decile) + real CCD-funnel prospects who inquired but never closed (teal — brighter = had appointment). The teal layer is THE callable warm-list, not random scored addresses.
Open map ↓Top-100 prospects score 1.00 vs 0.073 median = 13.7× lift. Driving features: neighbor density within 2mi (#1), recent Zillow activity, mid-range home value ($300-500K), 1950-1990 builds, central-CT cities. Square footage and beds are tertiary.
See findings ↓Topic model on items_summary discovered 8 archetypes. Two distinct $6K+ premium tiers worth productizing as a Concierge SKU. 75% of pre-1925 homes need Custom Angled.
See archetypes ↓Pricing-residual analysis BUT paired with collection rate. Adam Valendra books above predicted ($614/sale) and has the team's worst bad-debt rate (38% vs Jacob's 3%). Premium tickets you can't collect aren't premium — they're A/R. See Kyle Era for full pre/post.
See table ↓Cox proportional hazards model gives per-prospect P(buy within 12/24/60 months) given current ownership tenure. Top hazard-12mo: Glastonbury homes owned 20+ years (statistically "due"). Use for postcard timing.
See on map ↓Apples-to-apples comparison: the 21 months before Kyle bought in (Nov 2022 – Jul 2024) vs the 21 months since he closed (Aug 2024 – May 2026). Same length, same business, different leadership era.
vs flat baseline (the honest counterfactual), Kyle's tenure attributable lift is +$1,692,455.
Higher volume AND slightly higher mean ticket (+3% to $4,341) — productive trade-up.
Internet+Referral was 70%+ of revenue. Now 5 channels >5% each. Facebook 0.8 → 8.4%. Resilience built.
Southport ($41K), South Glastonbury ($44K), Sherman, Salisbury, Bozrah, Washington Depot...
| Source | Pre share | Kyle era | Δ pp | $ shift |
|---|---|---|---|---|
| 0.8% | 8.4% | +7.7 | +$587K | |
| Direct Advantage | 4.9% | 11.8% | +6.9 | +$601K |
| TV | 4.0% | 9.5% | +5.5 | +$482K |
| Finer Living | 0.2% | 4.7% | +4.6 | +$344K |
| Previous Customer | 4.0% | 6.2% | +2.2 | +$232K |
| 1. Internet | 39.7% | 36.7% | -3.0 | +$445K |
| Referral | 8.0% | 5.0% | -3.0 | -$90K |
| Rep | Pre close % | Post close % | Pre ticket | Post ticket | Δ ticket |
|---|---|---|---|---|---|
| Tim Choomack | 37.9% | 30.2% | $5,230 | $5,937 | +13.5% |
| Jacob Laskosky | 33.1% | 29.0% | $4,153 | $4,753 | +14.5% |
| James Willis | 34.4% | 33.5% | $3,854 | $4,389 | +13.9% |
| Stephen Clark | 48.1% | 41.3% | $4,448 | $4,716 | +6.0% |
| Robert Clark | 34.9% | 35.4% | $4,199 | $4,295 | +2.3% |
| Kevin Cerulo | 42.2% | 43.2% | $4,743 | $4,721 | -0.5% |
| Keith Rongey | 35.5% | 29.8% | $4,871 | $4,316 | -11.4% |
Pattern: top reps traded a few points of close rate for materially higher tickets. That's quality-of-deals discipline — and it shows up in the +29% revenue without much volume growth.
| Category | Pre % | Kyle era % | Δ pp |
|---|---|---|---|
| Foundation (high ticket) | 14.2% | 29.0% | +14.8 |
| Siding | 2.5% | 4.4% | +1.9 |
| Concrete | 0.7% | 1.3% | +0.6 |
| Custom Door | 26.7% | 25.4% | -1.3 |
| Powdercoat | 26.3% | 22.7% | -3.6 |
| BILCO (commodity) | 11.0% | 1.5% | -9.5 |
Foundation work doubled as a % of contracts. BILCO (commodity hatchway) cratered. Translation: CCD moved up-market on what it sells.
Bad-debt rose ($130K → $1.03M outstanding). Some of this is measurement artifact (post-Kyle contracts haven't had time to settle), but Adam Valendra's 38% bad-debt rate alone explains $246K of it. Worth a closer pre/post analysis of structural-vs-still-being-collected.
Two pools, both ranked by the live conversion model (LightGBM, AUC 0.733, calibrated). Each address shows lift signals — comparisons to typical CCD buyers, town-level historicals, lead-source close rates, and rank in pool. Click any row to expand the per-feature WHY breakdown (top 6 model contributions).
We scraped Zillow for 600 addresses next door to the densest CCD streets, then ran them through the v0.733 model. Median score: 52.8% — that's 7× the random baseline (7.3%). Every row here is an address that the model thinks looks like a CCD buyer AND sits near existing installs. Tightest possible door-knock target list.
Binary classifier trained on past-3-year CCD customers vs sampled prospects, using Zillow features (Zestimate, sqft, beds/baths, year built, school rating, tax rate, lot value, days-on-Zillow) plus CT property roll fields and city dummies. Then applied to all 768,036 CT residential addresses for ranking.
Baseline 0.703 → neighbor features 0.722 → Optuna tuning 0.727 → isotonic calibration 0.724. Final training on 14,946 enriched addresses (overnight scrape completion brought the dataset from 10K → 13.7K, +30%). AUC graduated from 0.635 → 0.677 → 0.724 in three training passes.
Calibrated top-scorers reach 1.00 probability (model says "if you call this address, expect a customer conversation"). Median pool score: 0.073. Top vs median lift: 13.7×.
6,476 customers + 7,265 non-customers (47/53 balance). Zillow + CT property + engineered + BallTree neighbor density at 0.5/1/2/5 mi (leakage-safe).
Every residential address in CT now has a calibrated conversion probability. Avg Precision 0.628. Top 1K + Top 5K CSVs downloadable below.
| Rank | Address | City | Owner | Score |
|---|---|---|---|---|
| 1 | 70 Southbury Rd | Roxbury | Pool, Michele J | 1.00 |
| 2 | 15 Iris Way | Berlin | Cwirka, Sharon & Brian | 1.00 |
| 3 | 432 Quassapaug Rd | Woodbury | Crumb, Erica L | 1.00 |
| 4 | 23 Nod Hill Rd | Wilton | Cromwell, Julie | 1.00 |
| 5 | 30 Nutmeg Dr | Somers | Cervone, Francesco & Laurie | 1.00 |
| 6 | 32 June Rd | New Milford | Escobar, Jose R + Cathy | 1.00 |
| 7 | 18 Ellsworth Lane | Canton | Nastri, David C | 1.00 |
| 8 | 16 Foxbriar Lane | Rocky Hill | McMannon, Judith TR | 1.00 |
| 9 | 6 Downing Dr | Preston | Lennon, Scott F & Melissa M | 1.00 |
| 10 | 6 St James Pl | Putnam | Lajoie, Lori E & Brian M | 1.00 |
| 11 | 16 April Way | Bloomfield | McLaren, Rochelle | 1.00 |
| 12 | 6 Vista Way | Bloomfield | Block, Karolina | 1.00 |
| 13 | 20 Herindeen Landing | Woodstock | Bocciarelli, Catherine | 1.00 |
| 14 | 30 Hodge Rd | Marlborough | Sena, Frances | 1.00 |
Calibrated probabilities — when the model says 1.00, it's stating "every comparable training example from this neighborhood-and-feature combination ended up being a CCD customer." Owner names from CT property roll. Pattern: 2003-2010 builds, 2000-3000 sqft, central + eastern CT mid-sized markets. Neighbor-density at 2mi has 0.605 univariate AUC — the spatial signal is real.
Final retrain on 14,946 enriched addresses (scrape completed 5/13 22:10 PM ET, retrain at 22:16 PM). Three training passes total: 0.635 → 0.677 → 0.724 AUC.
Gain-importance ranking from the final model — which features the model leans on most when predicting whether a CT homeowner becomes a CCD customer. The top three account for ~60% of total importance.
| # | Feature | Plain English |
|---|---|---|
| 1 | neighbor_count_2mi | Number of existing CCD customers within 2 miles. Strongest single signal — univariate AUC 0.605. The "word-of-mouth lift" we proved in early analysis (1.62× within 2mi). The model says: where CCD has already won, CCD will win more. |
| 2 | days_on_zillow | Recent Zillow activity on the listing. Higher page views = owner attention on their property. Especially: homes listed and then pulled (stale listings) — the "we tried to sell and decided to renovate instead" signal. 10× repeat rate vs never-listed homes. |
| 3 | zestimate | Home value sweet spot. The model favors mid-range homes ($300K-$500K) — not too cheap (deferred maintenance, owner can't afford CCD) and not too premium ($1M+ uses different contractors). Right-fit territory. |
| 4 | year_built / ayb | Older homes need more. Builds in the 1950–1990 band over-index. Aging bulkheads, original Bilco doors due for replacement, foundation work. Pre-2000 builds account for 75% of customer base. |
| 5 | zestimate_per_sqft | Price-per-sqft as quality proxy. Separates "premium-condition mid-size" from "cheaper square-footage". The renovated 2,000-sqft Bloomfield ranch beats the unrenovated 3,000-sqft fixer at the same total price. |
| 6 | tax_assessed / appraised_total | Town's view of the property. Independent of Zestimate. Correlates with deferred maintenance — homes assessed below Zestimate are often older but recently-improved (good signal); assessed above Zestimate suggests stale or over-valued listings. |
| 7 | sqft / living_area | Bigger basement = more bulkhead surface. 2,000-3,000 sqft homes over-index for foundation + waterproofing work. Sub-1,500 sqft skews to smaller items (egress windows only). |
| 8 | city = Berlin / Bloomfield / Rocky Hill / East Hartford | The historical CCD strongholds. Central CT corridor where neighbor density compounds. These cities collectively contribute ~20% of model importance via one-hot features. |
| 9 | schools_avg / schools_max | Family stability indicator. Good-school areas mean owners stay longer, invest more in property. Less mobility = more accumulated maintenance need over time. |
| 10 | tax_rate | Municipal cost-of-ownership signal. Secondary to assessed value. Higher-tax towns have older infrastructure, owner mix that has lived through maintenance cycles already. |
CCD's next customer is most likely a homeowner in central CT (Berlin / Bloomfield / Rocky Hill / East Hartford) within 2 miles of at least 5 existing CCD installs, in a 1950-1990 home with mid-range value (~$300-500K), recently active on Zillow (looking at their property, maybe listed and pulled). Word-of-mouth from existing customers is the strongest signal — far more than any individual property feature.
The strategic takeaway: CCD's job isn't "find weird outlier homes." It's double down on the neighborhoods where they're already winning. Geographic concentration > geographic diversification.
Facebook went from 18 → 232 inquiries/yr over the period. Close rate climbed 19% → 28%. Mean ticket $2.4K → $4.7K. Revenue per inquiry +88%. Clipper Print Ad is dying (volume -50%, close -57%).
560 contracts have outstanding balance > $100. Adam Valendra alone: 38.1% bad-debt rate, $246,765 outstanding (vs Jacob Laskosky 2.9%, Tim Choomack 5.5%). Immediate quality-of-contracts review needed.
$5K+ lifetime customers, no contact 5+ years, all have phone or email. Combined historical revenue $3,947,672. Even 5% reactivation = ~$200K in fresh contracts. List exported as CSV (button above).
Kevin Cerulo over-indexes on Waterproof by +394% ($14,219 mean ticket). Robert Clark on Foundation (+53%). Marc Benavides on Powdercoat (+133%). Ferdinando Crudele on Siding (+363%). Tom Giannotti on Concrete/Stair (+293%). Route inbound calls by primary product mention.
25,207 homes, only 0.18% CCD penetration, median home value $1.59M. Stamford 57K homes 0.30% pen at $585K homes. Fairfield already 1% but $704K. Volume whitespace: Bridgeport / New Haven / Hartford (50K+ homes each, <0.21% pen).
Only 349 of 7,827 customers (4.5%) ever repeat in 21 years. Median time to repeat: 234 days. Best predicted 5-yr repeat: ~7%. Don't model CCD as LTV-amortized — acquisition has to clear on contract #1.
First-job under $1.5K → 16.3% repeat rate. $3-5K first job → 3.5%. Theory: small first job is the customer's "let me test these guys" → big job later. Don't dismiss small inquiries.
@snet → 49.2% close rate, @optonline 44.7%, @comcast 44.2%, @aol 38.6%. National defaults (@gmail 34.9%, @yahoo 35.6%) sit lower. Long-term residents convert. @me.com customers spend the most ($5,535 ticket).
998 mover families (same name, multiple addresses) — 63 bought at both old + new house ($604K mover-repeat revenue). 288 had 2024+ activity → actionable warm-lead list right now.
Close rate jumps 28% → 31% around lunch. Thursday is the highest-revenue day. Best months: July ($1,306/inq) and October ($1,296). Worst: January ($941), December ($945). Schedule push-campaigns Thu 11am-2pm in summer/fall.
90-day window ending May 2026: $230,652 revenue. That's 4.07× his personal average. He's currently in the best stretch of his CCD career. Don't disrupt his flow. (Paul Rosenbeck's 2016 streak was 8.0× — singular career peak.)
28.5% of Internet-acquired customers return as "Previous Customer" — highest of any channel. 51.5% come back via Internet again. The digital-acquired customer is the most repeat-prone. 60% of Referral-acquired return via Referral — referrers create referrers.
82-89% of his jobs run over 90 days. Marc Benavides 98d median, 55% over. Kevin Cerulo (high-volume): 72d, only 35% over. Note: Tom is the Concrete/Stair specialist — concrete is structurally slow, so this needs separate scope-adjusted SLAs.
Inquiry-taken-by Leslie: 1,385 contracts, $6.6M revenue. Earline $3.3M, Jessica $3.2M, Eric Shearn $2.2M. Office staff are revenue-attributable. Pay her like a senior asset — she is one.
Cross-sell Markov: bilco → door (59%) is the dominant path. Stair → hatchway (64%). Doors → stairs (46%) or hatchway (31%). Egress → window well (100% — code-driven). Build the email playbook around these sequences.
Customers whose homes were on Zillow 90-365 days then pulled: 2.6% repeat rate vs 0.2% for never-listed. Small sample (n=76) but huge directional lift. Theory: failed listing → "improve what we have" → CCD repeat. Build a Zillow-stale-listing alerting layer.
Out of 17K contracts, only 5 have both quoted price AND final amount filled in. This single field, if mandated at appointment-set, unlocks pricing intelligence we've never had. Other CRM gaps: telemarketer/canvasser empty post-2020; "Tom Giannotti" vs "Tom Gianotti"; "Direct Advantage" vs "Direct Advantange".
Top fail reasons: Spouse Unavailable (1,046), Speak with Family Member (434), Office Quoting (175). 2,000+ failed appts where the decider just wasn't there → solid "both spouses confirmed" playbook hook. Price Objection (712) + Getting Other Estimates (537) = 1,200+ price-sensitive losses needing financing-options pitch.
2025 closed at $4.41M (919 contracts, $4,798 avg ticket). 11.2% compound annual growth since 2014. Linear projection: 2026 $4.40M, 2027 $4.68M. Exponential: 2026 $4.78M, 2027 $5.31M. The team's all-time peak was last year — still climbing.
Channel ROI (rev/inq ÷ CPI): Referral 458×, Internet 45×, Google Ads 42×, TV 36×, Facebook 34×. Suggested $100K split: $67K Internet, $17K Direct Advantage, rest spread. Counterintuitive: Facebook has lowest ROI (close rate 17% vs 33-43% elsewhere) despite highest growth. Long-term play vs short-term ROI tradeoff.
Same-day appt: 33% close. 91-180 day lag: 45% close. Best $/inq: 31-90 day lag ($1,776). CCD's product is a research purchase, not impulse. Don't kill yourself for same-day callbacks — send education content, follow up at week 4-8. The "instant follow-up = win" rule from B2B doesn't apply here.
Customers with non-CT area codes spend $4,592 mean ticket vs $4,207 for CT-area-code phones. 54% of OOS-phone customers come from Internet (they're researching online before moving). Concentrated in Fairfield, Stamford, West Hartford, Norwalk, Westport — premium suburbs. Out-of-state area code → auto-flag premium lead.
318 mega-clusters: same last name, same town, 3+ different first names, 3+ addresses. $776,148 combined revenue. Biggest: Williams family in Bloomfield (8 addresses, 8 people). Smith in West Haven (5/5), Miller in Hamden (6/6). Tactical: when CCD closes a contract, ask "anyone else in your family in town?" Cold-call siblings/parents/cousins.
Moran's I at customer-revenue level: 0.005 (effectively random). Earlier finding (1.62× lift within 2 miles) was about whether someone becomes a customer at all — that effect is real. But once someone IS a customer, how much they spend has zero relationship to neighbors' spend. Referral marketing → get the names; don't over-engineer cross-customer spend logic.
Greenwich (premium, senior reps, @me.com lookalike), Stamford (repeat-buyer reactivation, FB-first), Bridgeport (volume, pre-1970 angles, @snet email), New Haven (East Rock vs Fair Haven split), Hartford (greenfield, launch pricing, AM radio). Each deck: positioning + TAM math + talking points + channel rec + objection counter. Download button up top.
YTD through May 12: $1,152,770 in 213 contracts at $5,412 mean ticket — the highest mean ticket on record (up 13.5% YoY). +19.6% over 4-year avg for the same window. Projected full year: $4.71M. Don't slow marketing now — momentum is real.
21 years of city-level data reveals strong preferences. Lead with custom doors in New Haven (+18.6pp), Stamford (+13pp), Norwalk (+8.9pp). Foundation in Cheshire (+10.9) and Hamden (+10.6). Powdercoat (premium) in Enfield/Fairfield/Trumbull. Standard BILCO in Manchester/Glastonbury. Each rep's opening question should match the city.
Found 154 recent (2023+) failed appts where the source × salesperson combo normally closes 53%+. $322,000 in revenue potential. Biggest pocket: TV × Kevin Cerulo (20 missed, $108K potential). Reasons: ~half are "Decision-Maker Absent" — directly addressable by the failure playbook. Callable list: download button.
Top-quartile-revenue + paid-in-full customers cluster on: Fairfield County coastal suburbs (Fairfield/Milford/Stratford/Trumbull/Stamford), Internet or Referral acquisition, @gmail/@aol email, handled by veterans (Marc Benavides, Tim Choomack, Kevin Cerulo, Jacob Laskosky), $5K+ ticket. Use this list as the seed for Facebook lookalike audiences — not the full customer base.
Bridgeport: FB closes 10.3% vs 14% overall. Bloomfield: 8.7% vs 17%. New Haven: 14% vs 18%. Mean ticket $4,732 (above company avg) but per-lead intent is lower than other channels. Reconciliation: Facebook is the volume/growth play AND a top-of-funnel brand-building investment. Per-lead conversion needs work — better targeting + lead scoring + qualification before booking the appt.
Kevin Cerulo -78%, Robert Clark -26%, James Willis -24%, Jacob Laskosky -17% on Powdercoat over-index. These are the volume reps. Each leaves ~$150-200K/year of high-margin upsell on the table. Across the 4 top reps: $600-800K/year in unrealized attach revenue. Marc Benavides (+133%) shows what's possible. The fix is mandatory mention in every in-home pitch.
Bill D'Amico (Plainville, contractor email), Bruce Maneeley (S. Windsor, family-business), Haynes Construction (New Haven), Scalzo Property... These aren't homeowner reopens. They're trade-account pipeline reactivations — different motion entirely. Trade pricing, referral programs, monthly job-availability emails. Each successful reopen worth $15K-50K/year in recurring referral revenue. Split the reactivation list into consumer + trade tracks immediately.
From the dormant outreach analysis: (1) Bill D'Amico, Plainville — $26K spend, 5.1y dormant, contractor email, referral pipeline play. (2) Bruce Maneeley, S. Windsor — 3 contracts, "Prev Customer" source, family-business B2B = highest-confidence reopen. (3) Uwe Meier, Trumbull — 2 contracts, sbcglobal CT-local email, primed for the third purchase. 30 personalized call+email scripts await — download button above.
1,022 contracts (volume king) at 2.9% bad-debt rate (best on team) with +105% BILCO specialty (CCD's #1 SKU). Combination unmatched. Make him BILCO Product Owner — let him train the team on the company's most-attached SKU. Cross-team lift potential. Coaching plans for all 10 reps in dedicated download.
Don't ask "do you want to upgrade to powdercoat?" Instead, in the first 60 seconds: "What color is your trim?" Color becomes the default path, galvanized becomes the opt-out. Bundle pricing: "$4,720 with bronze, $4,140 galvanized — color's $580 of that." $580 lives inside a 12%-of-ticket frame, not standalone. Realistic Week-4 attach for Kevin: 8% → 22% = +$21K/yr just from Kevin.
Cross-sell Markov + recency decay → ranked callable list of existing customers most likely to buy soon. Top 25 are all 2026 customers within 1-3 months. 43 predicted to want Powdercoat (perfect timing for the new playbook), 157 "Other" cross-sells. Sum of 12-month expected revenue: $374,055. CSV downloadable up top.
Contracts with ≥95% balance owed — customer signed but never paid. Adam: 19 ghosts, $116,732 unrecovered. Kevin Cerulo 16 ($89K), James Willis 11 ($51K), Marc Benavides 9 ($46K). Internet leads have $119K of total ghost-revenue (highest source). Facebook $55K. Tighten deposit policy before installation, especially on Internet/Facebook closes.
Tim Choomack peak at 5-10y ($1,651/appt). Jacob Laskosky peak 5-10y ($1,903). Robert Clark steady climber, peaks 3-5y ($1,516). Keith Rongey peaks IN HIS FIRST 6 MONTHS ($1,960/appt) then declines — coachable signal (early intensity fading?). James Willis early-bloom 1-2y ($1,609). Ferdinando Crudele <6 months — too early to read. Use this for hiring + ramp expectations.
Cohort retention by acquisition year: 2015 cohort 1.8% by yr 1, 6.4% by yr 10. 2016 cohort: 2.2% yr 1, 5.7% yr 5. The repeat curve doesn't plateau — old customers DO trickle back, slowly. Long-tail retention is real but takes 5-10 years to materialize. Worth maintaining a nurture cadence for old customers (email at year 5, year 10) — costs nothing, captures the trickle.
Baseline 2026: $4.71M (best year ever).
Execute the 6 top recommendations at 50% effectiveness: dormant reactivation (+$130K expected), near-miss recovery (+$129K), powdercoat fix (+$55K), Adam A/R (+$50K), city routing (+$27K), next-best-product (+$50K) = +$441K expected.
Probability-weighted year-end: $5.15M. Best case (full execution): $5.59M. To hit $5M needs only 32.6% execution success across all levers — we modeled 50% as the base case. The action plan is the difference between "best year ever" and "$5M+."
For the first time, we can see what happens BEFORE the contract. 34,777 inquiries over 21 years. 88% became appointments (30,679). 22% converted to contracts (7,587 of inquiries; 25% appt→contract). Top fail mode is "decision-maker absent" — 34.6% of coded failures, recoverable with the playbook on the portal. Below: where revenue per inquiry diverges wildly by lead source.
| Source | Inq | Close % | Mean Ticket | $/Inq |
|---|---|---|---|---|
| Loading… | ||||
| Why it died | N | % of failed |
|---|---|---|
| Loading… | ||
Print and TV used to be everything. Internet now dominates. The story of marketing budgets you should be having in 2026 is whether the channels you're still spending on are the ones converting.
Heatmap of inquiry → close rate across the top 10 sources × every year. Look for cells that went from gold to gray — those are channels where lead quality is dropping even if volume holds. (Cells with <5 inquiries are blank.)
P25-P75 box per rep (inquiry → contract days). Fast medians = momentum. Wide boxes = inconsistent. A rep with 14-day median and tight box is a closer; a rep with 30-day median and wide box is leaving deals stuck in the pipeline.
Direct routing intelligence. If Kevin Cerulo closes Facebook at 38% but only 12% on Internet, route Facebook leads to him. Gray cells mean <10 inquiries, ignore those — small sample.
349 of CCD's 7,827 unique customer addresses (4.5%) have purchased multiple times. The #1 VIP — P.O. Box 385, Stratford — bought 7 cellar doors over 10 years for $129,867 lifetime — is almost certainly a property manager or landlord. Should be on a VIP nurture list, not in the general direct-mail queue. The Cox model shows median time-to-repeat is 234 days — nurture window opens ~6-8 months post-install.
| # | Address | Town | Jobs | Years | Lifetime $ |
|---|---|---|---|---|---|
| Loading… | |||||
Each row is a "cohort" — customers whose first CCD purchase was that year. Each column is "years since first purchase." Cell shows what % of that cohort bought again in year N. The Cox model said median time-to-repeat is 234 days — the heatmap shows whether that holds across cohorts and which years had unusually high repeat behavior.
4,740 customer addresses (red/gold by decile — colored by lifetime revenue) + 19,842 CCD-funnel prospects (teal — these are addresses that inquired or had an appointment but didn't close; brighter teal = had appointment, darker = inquiry-only). These are REAL CCD-system prospects, not random scored addresses. Click any dot for source + last-contact details.
The dots above show CCD's existing customers and the CRM's known prospects. The dots below are different — they are cold non-customer addresses the lookalike model thinks should buy, scored against the 21-year customer corpus. Every pin carries a full Zillow property panel and an expected-revenue number calibrated to CCD's historical close rates by tier.
| # | Tier | City | Address | Close % | Uplift | Exp $ | Zestimate | SqFt | Built | Near (mi) | Neighbors |
|---|
Loading cluster stats…
Loading…
The maps above show the proximity hypothesis visually. This is the same story in numbers. Each bar is the historical close rate (inquiry → contract) for addresses with that many CCD installs already within 2mi of them. The jump from "0 neighbors" to "1-2 neighbors" is the entire signal — most of the lift comes from the first few. Beyond that, density saturates.
The new ML model on 14,946 enriched addresses rank-orders all 768K CT residentials. Top-100 lift over median: 13.7× (calibrated probability). Old propensity v1 lift was 3.5× — adding neighbor density features + Optuna tuning quadrupled the signal at the top of the list.
Addresses within 2 miles of a prior CCD install hit top-decile 1.62× more often than random. We now bake this directly into the ML model as neighbor_count_2mi — it's the strongest single feature (univariate AUC 0.605), and the model uses 0.5/1/2/5-mi rings via BallTree.
Coastal corridor (Norwalk/Fairfield/Greenwich): highest mean ticket ($5K-$6K), Greenwich still 99% whitespace at $1.6M median home. Central-CT cluster (Berlin/Bloomfield/Rocky Hill/East Hartford): highest customer DENSITY, the ML model's favorite hunting ground for new conversions. Two strategies, two playbooks.
Of 7,827 unique customer addresses, only 349 (4.5%) ever repeat — but the cohort retention curve keeps rising past year 10 (6.4% by yr 10). Median time-to-second-contract: 234 days. CCD is structurally a first-job-payback business, not LTV-amortized.
Re-trained on 14.9K Zillow-enriched data (was 150 samples). Adam closes $1,851 above predicted per sale (mean). Tom Giannotti +$1,498. Tim Choomack +$1,334. Marc Benavides +$1,063. BUT: Adam's contracts have a 38% bad-debt rate ($246K outstanding) — Jacob L. sits at 2.9% on a much higher volume. Premium pricing you can't collect = A/R, not revenue.
Three-quarters of jobs at pre-1925 homes are Custom Angled doors. Pre-quote, pre-stock the truck, lead the sales call with it. The model agrees: year_built is the #4 feature by gain importance.
We ran Latent Dirichlet Allocation on the items_summary text of all 17,000+ historical contracts (21-year refresh). It found 8 natural clusters of job composition — including two distinct premium tiers ($6.3K and $6.5K mean) that look like productizable SKUs. The Foundation cluster doubled as % of contracts in the Kyle Era (14% → 29%).
| # | Address | Town | Ticket | Archetype | Reasons | Tier |
|---|---|---|---|---|---|---|
| Loading… | ||||||
A "fair-ticket" model predicts each job's expected price from address + neighborhood + job features. Comparing actual ticket to predicted ticket surfaces who's pricing high (under-priced segments) vs over-discounting.
| Salesperson | N sales | Mean residual | Mean ticket |
|---|
| Town | N | Mean residual |
|---|
Every cellar door sold, plotted by ticket size. The vertical lines mark the deciles. Median is $3,435 — that's where most of the business lives. Top decile starts around $6K; the rare $10K+ jobs are the premium tier that the LDA archetype model surfaced. Anything below the p10 line is suspicious — could be miscoded or fire-sale pricing.
21-year volume pattern: peaks July ($1,306 rev/inq) and October ($1,296). Trough: January ($941) and December ($945). Best closing day: Thursday. Best closing hour: 1 PM — close rate jumps 28% → 31% around lunch. Best months to fund pipeline-building: Dec-Feb for the spring spike.
Search any of 768,036 scored CT addresses. Returns ML v2 conversion probability (AUC 0.733, calibrated) + reason codes (neighbor density, year built, home value). Top scorers cap at 1.00; median across the pool is 0.073.
| Source | Inquiries Now | Inquiries Prior | Δ% | Close% Now | Close% Prior | Δpp | Revenue Now | Δ% Rev |
|---|