FT-Transformer · AI/ML Neural Pricing Architecture · production · 9-step pipeline · room-category level
Production · 9 Steps · Room-Category Level

Nine steps. Every room. Trained by the algorithm itself.

PriceCortex™ is the first AI/ML revenue management product that prices at the room-category level for US hospitality. Every recommendation walks through nine transparent steps — every signal versioned, every config row referenced, every decision captured as structured training data for the FT-Transformer neural pricing architecture.

9
Pipeline Steps
Room-Category
Pricing Level
2× / day
Comp Scrapes
100%
Decisions Audited
0 1 2 3 4 5 6 7 8 Context Match Base rank + base_metric Target Pressure Pace Adjust Constraints Persist · Train → FT-TRANSFORMER TRAINING SET
Live Competitive Intelligence

Real comps. Real rankings.
Real prices.

Here's one property's live compset — Hotel Indigo Frisco. Eleven ranked competitors across three priority tiers, updated twice daily via BDC scrapes, auto-classified by room category. This is the input that feeds Step 1 (Match) and Step 2 (Base Price) for every recommendation this property gets.

US_75034_001 MDM HOTEL GROUP
Hotel Indigo Frisco by IHG
IHG · Hotel Indigo Frisco, TX 132 rooms Upscale Boutique
14
Ranked Comps
4 / 4
Rank Tiers
42.8h
Scrape Age
Rank 1 · 3 comps closest true competitors · primary base-price driver
Drury Inn & Suites Dallas Frisco
$12442.8h
Hilton Garden Inn Frisco
$109570.8h
Sheraton Hotel Stonebriar
$9942.8h
Rank 2 · 4 comps premium upcomp set · waterfall fallback when R1 sold out
Aloft by Marriott Frisco
$12342.8h
Canopy By Hilton Dallas Frisco Station
$12242.8h
Hyatt House Dallas / Frisco
$116570.8h
Hampton Inn & Suites Legacy Park-Frisco
$10542.8h
Rank 3 · 2 comps broader Frisco / Plano set · tertiary tier
AC Hotel by Marriott Dallas Frisco
$14842.8h
Embassy Suites by Hilton Dallas Frisco Hotel & Convention Center
$122570.8h
Rank 4 · 5 comps outer tier · last-resort fallback
Hyatt Regency Frisco-Dallas
$282570.8h
Autograph Collection Dallas Frisco
$245demo
Homewood Suites Dallas-Frisco
$224642.8h
Home2 Suites Dallas-Frisco
$126570.8h
Cambria Hotel Plano - Frisco
$12242.8h
Twice-daily BDC scrapes · auto-classified by room category (King, Two Queen, King Suite, Accessible) · feeds Step 1 (Match) → Step 2 (Base Price) for every recommendation across this property's 132 rooms.
The Problem

Hotels price blind.
We give them eyes — per room.

Traditional RMS compares your hotel's average rate to the market average. That's like comparing apples to fruit salad. We match room category to room category: your King Suite against their King Suite. At the competitor level, by priority rank, updated twice daily. Per-category supply intelligence, not portfolio averages.

No other RMS in the US does room-category-level competitive intelligence for midscale & economy hotels.
PriceCortex™ · 9-Step Pipeline · v3

Nine steps. Every brick.
Every signal versioned.

Every room type gets its own competitively-derived price. Not a property-wide "base rate." Each room category priced independently against the same category at ranked competitors. Nine transparent steps — every config row, every signal, every multiplier captured as structured training data for the FT-Transformer.

Step 00 live

Context Lookup

// the setup · runtime pulls every input from versioned config

Every run starts by hydrating the property's metadata, room category list, ranked comp set, target occupancy, season, and the full RM-editable config from six tables. The audit trail starts here — every downstream signal carries the config version that produced it.

Inputs · per property × stay-date 6 config sources · 1 metadata source
→pricing_algo
Rank tier · base_metric · risk capacity per property
→target_rules
Target derivation method & growth multiplier
→pressure_rank_adjustments_v3
Asymmetric drop/raise multipliers per bucket
→booking_curve_inputs
Expected pickup % at every (season × DoW × t)
→overriding_rules
Hard floor · hard ceiling · flat lock per category
→competitor_ranking
Per-property ranked comp set with category map
Step 01 core innovation

Category matching.

// real like-for-like supply — not multipliers, not averages

For each room type you sell (King, Two Queen, King Suite, Accessible), the engine finds the same category at every ranked competitor and pulls today's price. "What is Hyatt House charging for THEIR King tonight?" Not a property-rate multiplier — actual category-level competitive intelligence.

Live match · Hotel Indigo Frisco · KING · 2026-05-28 5 of 5 ranked comps returned a price · fresh match
KING
Hotel Indigo Frisco
$174
current
KING
Hyatt House Frisco
$189
RANK 1
KING
AC Hotel Frisco
$209
RANK 1
KING
Hilton Garden Inn
$159
RANK 2
KING
Embassy Suites Frisco
$178
RANK 2
KING
Hampton Inn Stonebriar
$142
RANK 3
Step 02 differentiator rank + base_metric

Base price — derived two ways.

// the two RM-editable knobs that produce every base

Two config dials derive the base. Rank selects which competitive tier participates (Rank 1 = closest true competitors). base_metric is the within-rank aggregation — when a tier has multiple comps, the engine takes Lowest, Mean, Median, or Highest. Every tier is computed for audit transparency; only the configured cell drives the base. This is the room-category-level competitive logic no other RMS in our segments offers.

base_metric matrix · KING · 2026-05-28 Config → Rank=1 · base_metric=Mean → BASE $199.00
Lowest
Mean
Median
Highest
Rank 1
2 comps · ACTIVE
$189.00
$199.00
$199.00
$209.00
Rank 2
2 comps
$159.00
$168.50
$168.50
$178.00
Rank 3
1 comp
$142.00
single
single
single
Rank 4
empty
—
—
—
—
BASE = Mean(Rank 1) = mean($189, $209) = $199.00 · every other cell remains in the audit JSONB for transparency
Step 03 target_rules

Target derivation.

// the bar pressure rank measures against

Sets the bar Pressure Rank compares CY performance against. The basis — Occupancy or RevPAR (ADR × Occ) — and the increment come from target_rules per property. The historical method takes the same-day-match last-year actual and grows it: target = LY × (1 + incremental) — e.g. $70 ADR × 70% × 1.20. Property rules win; the seasonal Default ladder fills the gaps; if ADR is unavailable it falls back to occupancy automatically.

Derivation · KING · Thursday · Spring method = LY actuals × 1.20
LY Same Event
65.8%
×
Growth Multiplier
1.20
=
Target Occupancy
79.0%
target_occupancy · used as denominator for Step 4 RevPAR composite delta
Step 04 the judgment signal

Pressure rank — the judgment.

// weighted trailing Occ/RevPAR composite → 5-bucket score

A weighted composite of 30 / 7 / 2-day deltas vs Target maps to a 5-bucket pressure score (Very Low → Very High). The basis (Occ or RevPAR) and weights come from pressure_rank — RMs tune sensitivity per property without code changes, and the engine falls back to occupancy when trailing ADR isn’t reliable.

Pressure read · KING · 2026-05-28 composite_delta = −12.4% → bucket 4 · HIGH
1
Very Low
2
Low
3
Medium
4
High
5
Very High
Signal WindowWeightΔ vs TargetWeighted
Last 30 days50%−15.2%−7.60
Last 7 days25%−10.8%−2.70
Last 2 days25%−8.3%−2.08
Daily PMS feed0%parked—
composite_delta = −7.60 + −2.70 + −2.08 + 0 = −12.4% → bucket 4 · HIGH · drop multiplier amplified
Step 05 forward signal

Pickup pace — forward signal.

// actual OTB vs booking-curve expected at this exact point in the funnel

Where Pressure Rank is the trailing read, Pickup Pace is the forward read. For every stay-date, compare actual on-the-books against the booking-curve expected at this exact (days-to-arrival × time-of-day) lookup. Behind pace ⇒ drop signal · Ahead ⇒ raise signal.

Pickup curve · KING · 14d-out · 2:00 PM EST Spring · Thursday · −8.2% behind pace
NOW · 14d 90 days out 14d stay date 100% 66% 33%
Expected · booking_curve_inputs · 24.8% of target by 14d-out
Actual OTB · 22.7% of target → −8.2% behind
Actual OTB
23 rooms
Expected OTB
25 rooms
Pacing Gap
−8.2%
Signal
BEHIND
Step 06 asymmetric multipliers

Price adjustment.

// pressure × pace combined · gap drives direction, bucket drives magnitude

The combined gap (Steps 4 + 5) sets direction — drop or raise. The pressure bucket sets magnitude via asymmetric multipliers from pressure_rank_adjustments_v3. Drop multipliers amplify when under-pacing (cut harder to recover demand); raise multipliers amplify when outperforming (capture upside). Capped by max_drop and max_gain.

multiplier matrix · pressure × side applied: bucket 4 · drop side · 1.20×
PR 1 · Very Low
PR 2 · Low
PR 3 · Medium
PR 4 · High
PR 5 · Very High
Drop sideamplifies under-pacing
0.80×
0.90×
1.00×
1.20×
1.30×
Raise sideamplifies outperformance
1.30×
1.10×
1.00×
0.85×
0.80×
BASE $199.00 + pressure-adjusted drop $20.00 × 1.20 = −$24 nominal · capped to −$20 → $179.00 · within max_drop $30 cap
Step 07 weather · events

Weather & events.

// demand environment → 1–5 score → premium or discount

Two market signals score the stay date 1–5: local weather (from weather_snapshots) and demand events (from the per-property events calendar — concerts, conventions, holidays). 1 suppresses demand → discount · 3 is neutral · 5 lifts it → premium. The combined score becomes a multiplier applied after the pace/pressure adjustment and before the hard constraints — so a sold-out-town weekend earns a premium and a washout midweek softens automatically.

Demand environment · KING · 2026-05-28 weather 4 · event 5 → score 5 · 1.10× premium
1
Discount
2
Soft
3
Neutral
4
Firm
5
Premium
$179.00 pace-adjusted × demand multiplier 1.10 (weather 4 + event 5) = $196.90 → passed to constraints
Step 08 overriding_rules

Constraints.

// hard floor · hard ceiling · flat lock — per-property veto

Per-property veto from overriding_rules — hard floor, hard ceiling, or flat-price lock. Breaches are logged with full reasoning. Owner is proactively notified: "Your ceiling fired 12 times last weekend — consider raising it."

Constraint check · KING · 2026-05-28 computed $196.90 · floor $145 ✓ · ceiling $245 ✓ · no breach
Floor
$145
Ceiling
$245
No clamping. Computed price falls comfortably between hard limits — recommendation publishes at $196.90.
Step 09 FT-Transformer hand-off

Persist · publish · train.

// every step, every signal, every config row → JSONB → neural training set

The final price is written to pricing_recommendations_v3 with the complete audit trail — every step, every signal value, every config_version_id referenced — captured as structured JSONB. This is the labeled dataset the FT-Transformer trains on. Every recommendation today contributes to the neural network that refines this engine tomorrow.

Output flow · recommendation → outcome → model audit JSONB · 9 step trails · ~12 KB/rec
$196.90
Persist
pricing_recommendations_v3 · 9-step JSONB
→
+30d
Outcome
join to actual ADR · occ · RevPAR per stay-date
→
∿
FT-Transformer
learns elasticity per scenario · refines decisions
Every recommendation is a labeled training row. The 9-step audit (config_version → signal_values → adjustment → final_price → outcome) is exactly the structured input the FT-Transformer consumes.
FT-Transformer · Neural Architecture

Every recommendation is a training row.
The model learns from every decision.

PriceCortex™ is built on FT-Transformer neural pricing architecture — the same family of feature-tokenizer transformer models that lead the field on tabular prediction. Every 9-step audit trail — config version, signal values, multiplier choice, final price — becomes a labeled training row. As decisions accumulate alongside actual outcomes, the model refines its read on price elasticity per (property, segment, scenario).

FT-Transformer · Architecture (abstract)
Production · every recommendation = a training row
INPUT · 9 STEP OUTPUTS FEATURE-TOKENIZER · ATTN TRANSFORMER STACK · ×N PREDICTION · OCC + ADR + REVPAR occ % ADR $ RevPAR
01
Tabular-native
FT-Transformer is purpose-built for structured tabular data — exactly the shape of competitive pricing signals. Outperforms generic deep nets on this class of problem.
02
Feature-tokenizer attention
Each of the 8 pipeline outputs becomes a token. Self-attention learns which signal matters most for each scenario — different mix for weekends vs weekdays, leisure vs corporate, peak vs trough.
03
Three prediction heads
Predicts occupancy, ADR, and RevPAR per price point — so the engine can optimize for whatever the owner prioritizes. Not a black box; the audit trail comes with every prediction.
AI Surfaces · Halo + Audit

How owners talk to the algorithm.

Every PriceCortex™ recommendation is reachable through conversational AI and a full 9-step audit drill-down. Halo™ is the owner-facing AI revenue manager — grounded in the same data the engine sees.

Halo™
AI Revenue Manager.
Owners ask: "Why is my Friday King at $179?" · "Should I drop my floor next weekend?" Halo answers with grounded data from your property — pressure rank, pickup pace, comp rates, history. Domain-expanded across RM concepts and hotel industry context.
9-Step Audit
Every decision — unpacked.
Click any brick in the RM view to open the full 9-step audit modal. Every step shows what happened and why — the same JSONB trail that becomes FT-Transformer training data downstream.
Decision Impact
Config impact dashboard.
The first user-visible learning surface. Every config tweak — rank, base_metric, pressure weights, multiplier values — gets a measured lift effect. RMs tune; PriceCortex shows the result.