Smart Forecast — hybrid forecasting

In lesson 20 we built a 3-layer forecast model by hand — last year’s base, pace extrapolation, manual corrections. In lesson 37 we added the booking curve: the date-level booking fingerprint. Now we assemble the two into a mature, hybrid forecasting approach — this is the internal logic of the Peaqplus Smart Forecast module.

“Hybrid” is the key word: we don’t run one model but several in parallel, and combine their outputs with context-sensitive weights. One of the oldest lessons of forecasting research is that a combined forecast is typically more accurate than any of its components alone — which is why every mature RMS forecast works this way today. At the expert level (lesson 55: Smart Forecast Enhanced — the longer-horizon hybrid model) we’ll see the extended version; this lesson walks through the classic, statistical hybrid, on real numbers.

What is a hybrid forecast?

A hybrid forecast is a framework: several different forecasting methods run in parallel for the same date, and the final output is a weighted average of the estimates.

The problem with the single-model approach is that every method has a blind spot:

• The historical average (comparable-date model) is strong on stable patterns but blind to this year’s specific situation — it can’t see when pace is breaking away from the usual path.
• Pace extrapolation works from this year’s actual booking tempo, but it is noisy early on (extrapolating from little OTB) and can overreact to a random spike.
• Neither knows what only the RM knows: the group contract signed yesterday, next week’s campaign, the freshly discovered concert.

The hybrid model runs all three information sources and weights them by context: on an event day the context layer gets a heavy weight; on an average Wednesday, pace dominates.

The three layers of Smart Forecast

The internal stack of Peaqplus Smart Forecast is built on three layers. Each produces its own estimate; at the end they are combined with weights. Let’s walk through one concrete day at Hotel Peaqplus City: Saturday, Dec 16 — the Christmas-market peak weekend, with an arena concert recently added to the event calendar for that evening. The forecast is produced 7 days before arrival.

Layer 1 — the comparable-date base

This is the “last year’s base” from lesson 20, automated: the model selects the comparable days itself from the last 3 years (day-of-week aligned, calendar-adjusted — a Christmas-market Saturday matched to a Christmas-market Saturday), and weights them by recency:

The weighted base: 0.5 × 88 + 0.3 × 85 + 0.2 × 80 = 85.5%.

Layer 1 forecast: 85.5%. Lesson 20’s manual model computed this layer as a simple average — the machine version goes further by calibrating the comparable selection and the year weights from data.

Layer 2 — pace extrapolation from the booking curve

This is lesson 37’s method: we measure the current OTB against the same point of the standard booking curve, and project the gap out to check-in.

Pace is clearly ahead of the curve, and the extrapolation projects a practically full house.

Layer 2 forecast: 98.8%.

(A subtlety from lesson 37: if the date has event history, the model would compare against an event curve, not the standard Saturday. Here the concert was only recently added to the calendar, with no event past — so the standard curve does the work, and Layer 3 brings in the context.)

Layer 3 — event and context corrections

This is not a statistical layer: the event calendar, the sales team’s knowledge and marketing activity go here — the things the number series can’t see yet.

Note the double-counting rule (introduced in lesson 20): whatever is already in the OTB is not added again in Layer 3. That’s why the concert correction is small (most of its demand is already in pace) and the group’s is large (it shows up nowhere yet).

(Whether a 10-room group should be accepted at all on an already-strong Saturday — that’s the displacement question of lesson 40.)

Weighting the three layers

The key question: how do three numbers become one forecast? The weights are context-dependent:

The mechanics of the combination: Layer 3 by itself is not a forecast but a set of corrections — so the model turns it into one: it starts from the average of the two statistical layers and applies the corrections on top.

• Layer 3 base: (85.5% + 98.8%) / 2 = 92.2%
• With corrections: 92.2% + 17.5 pp = 109.7% — physically impossible, so the model caps it at 100%.

A number running above 100% is not an error signal but information: the part above the ceiling is demand we cannot serve. In lesson 39 (unconstrained vs. constrained demand) we learn to measure exactly this above-100% demand — and turn it into a pricing decision.

The final, weighted forecast with event-day weights:

0.15 × 85.5 + 0.35 × 98.8 + 0.50 × 100 = 12.8 + 34.6 + 50.0 = 97.4% — that is, ~78 rooms out of 80.

The model attaches a confidence range: 97.4% ± 4 pp — reality will very likely close between 93% and a full house. This is the Smart Forecast output.

The same stack runs for ADR (comparable-ADR base, mix-adjusted pace layer, event ADR correction) and at segment level too — we walked the example through on occupancy.

Calibrating the model

The weights and layers are not carved in stone — the model continuously measures itself:

• Daily — actuals measurement. For every closed day, the forecast vs. actual gap is computed: the MAE / MAPE / bias measures known from lessons 19 and 26.
• Weekly — weight calibration. If the pace layer keeps beating the comparable-date layer, the module raises Layer 2’s weight — separately per context.
• Monthly — model-level tuning. Reviewing the comparable-selection logic, the curve averaging and the correction thresholds.
• Yearly — macro review. Peaqplus advisors and the RM team review together: which segments the model is strong on, where the weak spots are.

This cycle is the hybrid model’s real strength: not that it’s accurate on day one, but that it gets more accurate month after month.

What does it add over the Excel model?

In lesson 20 we saw what a manual, Excel-level forecast can do. Smart Forecast goes beyond it on six points:

1. Speed. The 30-day manual forecast takes 15-20 hours a month; Smart Forecast recomputes all 90 days in seconds — every night.
2. Segment-level precision. By hand you typically forecast at total level — there’s no time to build a model per segment. The machine works with a separate curve and pace per segment.
3. Continuous calibration. The manual model is static, refreshed monthly; the hybrid measures and adjusts itself daily.
4. Confidence range. A manual forecast gives one number; Smart Forecast gives a range (97.4% ± 4 pp). Useful in owner and bank communication too: a range is more honest than a falsely precise number.
5. Anomaly alerts. If a date’s forecast moves sharply from one day to the next, the system flags it — invisible in a manual model.
6. Multiple models, context weighting. Excel works with one method; the hybrid with three, and it knows when to listen to which.

The accuracy gap in numbers, based on industry experience: a well-calibrated hybrid RMS forecast typically runs around a 4-7% MAPE, while a manual, Excel-level forecast is usually in the 12-18% range. Lesson 26 showed what that means in the monthly control: a planning error half to a third the size.

The limits of the model

Smart Forecast is a classic statistical model — not magic, and not AI. Four limits worth knowing:

1. New patterns. If the hotel’s demand structure changes structurally (e.g. a new MICE segment builds up), the model needs 2-3 months of data to recalibrate — read it more cautiously until then.
2. Sudden shocks. No statistical model predicts a COVID-style macro break — the model only adjusts after the fact.
3. Compset moves. The model works from its own booking data; it doesn’t directly see a competitor’s dramatic rate cut — only once it shows up in our own pace. Compset knowledge is brought in by the RM — from lesson 32’s shopping routine, as Layer 3.
4. A suggestion is not a decision. Sometimes the RM’s knowledge overrides the model — which is exactly why you need to understand what happens under the hood. Smart Forecast makes a suggestion; the responsibility is yours.

Part of the answer to these limits comes at the expert level: lesson 55 (Smart Forecast Enhanced — the longer-horizon hybrid model) and lesson 56 (Pricing Engine — ML-based rate recommendations).

Smart Forecast in one day — Daniel’s routine

What does all this look like in practice, on an average Monday morning?

8:30 — reviewing the forecast refresh. The module recomputed all 90 days overnight. Daniel opens the next 30 days’ forecast vs. budget vs. actual grid, plus the “top forecast changes” list — where today’s forecast moved sharply against yesterday’s.

8:35 — confidence filtering. On narrow-range days (±2 pp) he trusts the model; wide-range days (±8 pp) get a manual look.

8:40 — the anomaly list. Five days flag today: 3 positive (pace ahead of the curve), 2 negative (slowing). A few minutes each: he accepts the forecast or applies a Layer 3 correction.

8:50 — actions. The anomaly days trigger rate revisions, restrictions or promos (lessons 35-36), and this forecast is also the input for the weekly revenue meeting (lesson 28).

Roughly 20 minutes — while lesson 20 showed the same job by hand is 15-20 hours of Excel work a month. And the difference isn’t just time: the machine computes every day, for every segment, with the same methodology — and the RM focuses on what genuinely needs a human.

Key takeaways

• A hybrid forecast = several methods in parallel, combined as a weighted average. The combined estimate is typically more accurate than any single model.
• Smart Forecast’s three layers: Layer 1 (comparable-date base), Layer 2 (pace extrapolation from the booking curve), Layer 3 (event and context corrections).
• The weighting is dynamic: on an average day pace dominates (60%), on an event day the context (50%), on a slow day the historical anchor (50%).
• Layer 3’s golden rule is avoiding double counting: whatever is already in the OTB does not go in again as a correction.
• The model calibrates continuously — daily actuals measurement, weekly weight adjustment, monthly tuning, yearly review.
• The hybrid model is statistical, not AI: it sees new patterns, sudden shocks and compset moves only with a delay. The RM’s knowledge and right to override are part of the system — not a flaw in it.