Insights & Suggestions · DET Activation Planner

Learning loop · 🇩🇪 Germany

Insights & calculated suggestions

Every data load is appended to a history store. A propensity model is retrained on campaign outcomes, trends and forecasts are read off the collected series, and the suggestions below combine both with the trigger rules. Simulated rows are labeled and are replaced as real captures accumulate.

History store · …

0 observations · 0 campaign outcomes · nothing captured yet

Set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY (schema in supabase/schema.sql) to collect into a shared Postgres warehouse instead of this browser.

Demand propensity · observed

33% real inputssimulated population

Target: monthly Dubai searches per eligible adult (Google TAC export). Ridge regression on 3 catchment features, 9 catchments · fit R² 0.50 · leave-one-out R² -0.17 · too few catchments to generalise yet; add UK and KSA exports or census data

Origin temperature (°C)+0.119simulated
Active-silvers index+0.075simulated
Search lookahead (days)-0.053export
Weight per standard deviation, in searches per adult. Provenance per feature; simulated inputs become open or export as sources load.

Observed	Predicted	Latent index	Propensity
0.30 · 500K/mo	0.60	50	78
0.71 · 700K/mo	0.32	222	68
0.45 · 500K/mo	0.29	155	63
0.39 · 400K/mo	0.32	124	56
0.26 · 200K/mo	0.46	58	53
0.19 · 100K/mo	0.33	56	51
0.13 · 70K/mo	0.26	51	39
0.22 · 400K/mo	0.09	249	24
0.03 · 20K/mo	0.14	20	17

Observed = searches per eligible adult per month. Predicted = leave-one-out, so each catchment is scored by a model that never saw it. Latent index above 100 = searching more than its profile predicts. Propensity = predicted intensity on a 0–100 scale where 0.6 per adult is 100.

Propensity model

untrained

Needs at least 8 campaign outcomes for Germany. Backfill simulated outcomes to see the mechanics, or connect partner booking feeds for real ones (0 so far).

Propensity by catchment · Gastronomy & Dining

Baseline	Model	Δ	Relevant	Expected bookers
3.6%	—	—	193,536	2,648
2.9%	—	—	259,740	2,583
2.8%	—	—	220,416	2,469
3.0%	—	—	142,296	1,765
3.0%	—	—	145,656	1,560
3.1%	—	—	114,912	1,339
2.7%	—	—	67,392	600
2.4%	—	—	77,832	530
2.5%	—	—	64,152	508

Baseline = 2.5% prior booking rate × affinity index. Model = learned bookings per reached individual. Expected bookers = relevant audience × partner share × rate.

Demand trend & 14-day forecast

Fewer than 14 daily points for Germany (0). Load Wikipedia pageviews or backfill to see a trend.

history (simulated)
live captures
Holt forecast

Calculated suggestions · Munich Catchment × Gastronomy & Dining

Open the flight by 2026-08-16 to catch the search peak

· Campaign opens 17 days before first departures; DE searches peak 70–77 days out
· Departures in the first weeks of the travel window are currently past their booking peak

Covers 60 more days of peak search for early departures

timing · high · 65

3 of 4 triggers are live — shift to a conversion objective

· Temperature differential / freeze trigger: Munich Catchment: 3°C Overcast, 2 rain days → ACTIVE
· Holiday departure window · Bavaria (Bundesland + national calendar): Herbstferien 2 Nov – 6 Nov 2026 · launch 17 Aug – 24 Aug 2026 ✗ outside campaign
· Search surge anomaly (Google TAC / Trends): ACTIVE — +33% vs 14-day avg

Move 30% of consideration budget into Partner Booking Referrals while the window is open

budget · medium · 62

Load open data to replace simulated weather, holidays and demand

· Open-Meteo, Wikipedia pageviews, Nager.Date and ECB rates need no keys
· Each load is appended to the history store, so trends improve with every visit

Live inputs for rules 1–4 and the first real observations for the model

data · high · 35

Model has 0 real outcomes — connect partner booking feeds

· 0 rows in training, 0 simulated
· Confidence moves to medium at 15 real outcomes and high at 40

Propensity estimates become market evidence rather than modeled priors

data · high · 30