Price Elasticity of Demand

The Model

What we measure

Competitor spread = Gen H's headline rate minus the best competitor's rate in our relevant competitive set. Measured in basis points (bps).

Negative spread = Gen H is cheaper (e.g. −20 bps means we're 0.20% below the next-best lender)
Positive spread = Gen H is more expensive (e.g. +140 bps at 90% LTV)
Continuous, not ordinal — being £1 cheaper and £50 cheaper are very different propositions for a broker

Competitive Sets

LTV Band	Competitive Set	Rationale
60–80%	Specialist Only (~40 lenders)	Gen H competes head-to-head with Kensington, Aldermore, Pepper, Precise
90%	Specialist + Tier 3 (~25 lenders)	Building societies serve overlapping customers at this LTV; including them nearly doubles explanatory power
95%	Specialist + Tier 3 (~30 lenders)	Gen H is the only Specialist — T3 building societies are the nearest cheaper alternative

Weekly regression

volume_week = β₀ + β_month + β_trend · t + β_weekdays · n + β_spread · competitor_spread + ε

Calendar controls absorb seasonality (house-buying season peaks in Sep, troughs in Dec/Jan), the structural growth trend, and short-week effects (bank holidays).

Appendix: How We Got Here

The competitor spread model is the result of six analytical phases over multiple sessions. Each phase tested a hypothesis, found its limits, and pointed to the next refinement. This appendix documents the journey and provides the evidence for each decision.

A1. Why Absolute Margin Is Wrong

Hypothesis: Gen H's margin over swap rates (our "price") should predict volume directly.
Verdict: Rejected. Absolute margin explains <3% of weekly volume variance.

We started by regressing daily volume against Gen H's margin over swaps (product rate minus Aston swap rate). The first-differenced correlation was r = −0.27 to −0.32 across pipeline stages — statistically significant but economically tiny (4–10% of variance).

Why it fails: Brokers don't think in absolute margin. They think "who's cheapest for my client?" A 10bps margin increase while still the cheapest lender has zero impact. A 5bps increase that drops you from 1st to 3rd is devastating. The same margin can mean completely different competitive positions depending on what competitors are doing.

A2. Why Full-Market Rank Is Meaningless

Hypothesis: Gen H's rank among all ~110 lenders (from Twenty7Tec) should predict volume.
Verdict: Rejected. Ranking includes Halifax, HSBC, Barclays — lenders whose customers would never consider Gen H.

Gen H's full-market position was rank 31–63 out of 110. This is meaningless for predicting volume because most of those 30+ cheaper lenders serve a completely different customer segment (prime, employed, clean credit). Rate spread to the full-market best was moderately predictive (r = −0.25 to −0.32) but noisy.

A3. How We Chose Competitive Sets

Method: Let the data decide. We tested 5 competitive set definitions empirically and picked whichever best predicted volume (highest ΔR²) at each LTV.

The Five definitions tested:

#	Definition	60% winner?	80% winner?	90% winner?
1	Specialist Only	Yes	Yes	No
2	Specialist + Tier 3	No	No	Yes
3	Specialist + Tier 2	No	No	No
4	Specialist + Tier 2 + Tier 3	No	No	No
5	All except High Street	No	No	No

At 60–80% LTV, Specialist Only wins because Gen H's actual competitors are Kensington, Aldermore, Pepper, and Precise. Adding Tier 2/3 building societies dilutes the signal with lenders serving different customer profiles.

At 90% LTV, adding Tier 3 nearly doubled DIP explanatory power (2.9% → 5.2% on the old rank metric). Building societies DO serve overlapping customers at 90% — they're the nearest cheaper alternative when a broker is weighing Gen H's criteria flexibility against the price premium.

A4. Why Weekly Beats Daily

Problem: Daily R² values of 3–6% looked unimpressive and understated the real effect.
Cause: Poisson counting noise dominates at small daily counts.

Daily volume at per-LTV level is small (e.g. mean 4.4 DIPs/day at 90%). For a Poisson process, 80% of daily variance is random arrival timing. Weekly aggregation fixes this: signal accumulates 5× while noise only grows √5 ≈ 2.2×.

A5. Why Spread Beats Rank

The key insight of this analysis. Ordinal rank collapses magnitude information. Being rank 1 by 1bp and rank 1 by 100bp look identical. Competitor spread preserves the full picture.

Head-to-head comparison (weekly ΔR² beyond calendar)

At 90% LTV, the improvement is dramatic: spread explains +18.9% vs rank's +14.7%. In the joint model (both rank and spread as predictors), spread remains significant (p=0.005) while rank becomes marginal (p=0.092) for DIPs. Spread subsumes rank — rank was just a noisy proxy for the continuous spread all along.

At 60–80% LTV, rank "wins" the head-to-head, but the mechanism is the same. Gen H oscillates around the cheapest position (spread crosses zero). The rank variable captures the threshold — "are we cheapest or not?" — which is just the sign of the spread. With more data or a nonlinear specification, spread would capture this too.

A6. The 95% LTV Criteria Moat

Finding: 95% LTV volume is not price-driven at conversion stages. Gen H has a criteria moat: customers apply because no cheaper lender will approve them.

With the Specialist-only competitive set, Gen H was the only lender at 95% — rank 1 of 1 for 352/366 days. Zero variation, zero signal. This led the previous analysis to write off 95% entirely.

With Spec+T3, there IS spread variation (85–196 bps). Spread is significant for Calculator sessions (ΔR² = +4.3%, p=0.006): when Gen H's premium narrows, more brokers check our rates. But by the DIP stage (p=0.13) and beyond, the signal disappears. Brokers browse based on price but apply based on criteria.

Executive Summary