Script

Tourism Demand Prediction Using a Multisource Temporal Fusion

Transformer with BERT-Based Sentiment Signals
Timing at a Glance
Target total speaking time: 9 minutes 45 seconds to 10 minutes.
Section Slides Target time
Opening and motivation Slides 1-3 1:25
Gap, dataset, preprocessing, and Slides 4-8 3:10
feature fusion
Sentiment model and forecasting Slides 9-10 1:20
setup
Main results and validation Slides 11-12 1:35
comparison
Ablation, explainability, and forecast Slides 13-15 1:55
demonstration
Discussion, limitations, and Slides 16-18 2:10
conclusion
References transition Slide 19 0:05

Presentation Style Guidance

• Speak in short phrases and pause briefly after each major result.
• Do not read every number in the tables. Mention only the strongest result and the interpretation.
• Use Slide 12 as a strength: it shows honest evaluation rather than cherry-picking.
• For Slide 19, simply acknowledge references and invite questions.
One-sentence central message
A Dubai-specific multisource framework improves short-term tourism forecasting when search behaviour and BERT-
based sentiment are combined with transparent validation.
Slide-by-Slide Speaker Script
Use this section as your main practice script. Each slide is written in a concise conference style and includes a
delivery cue.

Slide 1 - Title Target: 0:25

Good morning/afternoon, respected examiners and colleagues. My name is Muhammad Abdullah, and I am
presenting my dissertation titled Tourism Demand Prediction Using a Multisource Temporal Fusion Transformer with
BERT-Based Sentiment Signals. This study focuses on forecasting short-term tourism demand in Dubai by
combining Google Trends search behaviour, tourist review sentiment, and deep learning-based time-series
modelling.
Delivery cue: Open calmly, make eye contact, and do not rush the title.

Slide 2 - Agenda Target: 0:20

The presentation follows a simple structure. I will first introduce the motivation and research gap, then explain the
dataset and modelling framework. After that, I will present the main results, ablation and explainability findings, and
finally close with limitations, future work, and the conclusion.
Delivery cue: Keep this slide brief; it is only a roadmap.

Slide 3 - Research motivation and study aim Target: 0:40

Tourism demand forecasting supports practical decisions such as staffing, pricing, capacity planning, and destination
management. In Dubai, this task is challenging because demand changes with seasonality, events, online travel
interest, and external shocks. The aim of this research is to examine whether forecasts improve when historical
demand is combined with online search behaviour and BERT-based sentiment from tourist reviews.
Delivery cue: Emphasise the practical value: decisions, not just model accuracy.

Slide 4 - Research gap and contribution Target: 0:45

The literature shows value in deep learning, search trends, and online reviews, but these signals are often studied
separately. The gap addressed here is the lack of an integrated, explainable, multisource pipeline for destination-
level short-term forecasting. This study contributes a Temporal Fusion Transformer framework for Dubai, compares it
with ARIMA, Random Forest, XGBoost, and LSTM, and evaluates performance using both held-out and rolling-origin
validation.
Delivery cue: Use this slide to show originality and academic positioning.

Slide 5 - Dataset and multimodal study design Target: 0:35

The study uses two main data sources. The first is weekly Google Trends data for Visit Dubai, Dubai Hotels, and
Dubai Tourism. The second is a tourist review corpus containing 4,720 reviews, with 4,545 valid review dates. The
final weekly demand series contains 261 weeks, and after feature construction, 257 weeks are used for modelling.
Delivery cue: Mention numbers clearly; they show scale and credibility.

Slide 6 - Constructing the Dubai demand proxy Target: 0:35

Because official visitor arrivals were not used, a reproducible demand proxy was created from Google Trends. The
raw search series were cleaned, interpolated, normalised, and averaged into one composite demand index. This
proxy captures online tourism interest over time and preserves the seasonal pattern of Dubai, especially stronger
demand during the cooler months.
Delivery cue: Point to the chart and explain it as a cleaned signal, not raw data.

Slide 7 - Review corpus and sentiment signal Target: 0:35

The review corpus provides the text-based modality. It is strongly positive, with about 79.9 percent five-star reviews
and an average rating of around 4.62. Although this limits sentiment variation, the text still contains useful signals
about service quality, location, and price or value. These aspect-level signals are aggregated weekly and aligned
with demand.
Delivery cue: Acknowledge the positivity as both useful and limited.

Slide 8 - Feature engineering and multimodal fusion Target: 0:40

The final feature set combines demand lags, rolling statistics, seasonal indicators, Google Trends variables, review
volume, BERT polarity, and aspect sentiment. All features are aligned weekly. The fusion idea is straightforward:
temporal and search features are combined with sentiment features into one modelling vector, allowing the models to
learn from both structured and text-derived information.
Delivery cue: Make the fusion sound clear and practical rather than over-complex.

Slide 9 - DistilBERT sentiment module performance Target: 0:35

Before using sentiment in forecasting, the DistilBERT model was validated separately. It was trained on 3,776
reviews and validated on 944 reviews. The model achieved validation accuracy of 0.8369 and weighted F1 of 0.8225.
This confirms that the review text produced a reliable behavioural signal before being added to the forecasting
models.
Delivery cue: Stress that sentiment was validated before fusion.

Slide 10 - Forecasting models and validation protocol Target: 0:45

Five models were compared: ARIMA as the statistical baseline, Random Forest and XGBoost as machine-learning
baselines, LSTM as the recurrent deep model, and TFT as the proposed transformer model. Two validation
approaches were used: chronological held-out testing and rolling-origin validation. This is important because it
prevents the results from depending on only one favourable evaluation design.
Delivery cue: Pause briefly after naming the two validation methods.

Slide 11 - Held-out test results Target: 0:45

On the held-out test set, the Temporal Fusion Transformer achieved the strongest performance. It obtained MAE of
0.0176, RMSE of 0.0247, MAPE of 5.87 percent, and R squared of 0.9518. This shows that TFT was best at
capturing long-range seasonal structure and interactions between multiple data sources in the final test period.
Delivery cue: Focus on the conclusion, not every number in the table.

Slide 12 - Rolling-origin validation Target: 0:50

Rolling-origin validation gives a different result. In this setting, ARIMA performs best, followed by XGBoost and
Random Forest. This does not invalidate the TFT result; instead, it shows that different protocols measure different
forecasting abilities. ARIMA is strong for local short-window tracking, while TFT is stronger in the held-out non-
stationary test period. Therefore, the study reports both results honestly.
Delivery cue: This slide is important: present it as methodological maturity.

Slide 13 - Ablation study Target: 0:40

The ablation study tests whether sentiment actually helps. The result is not uniform across models. Sentiment has
almost no effect on ARIMA, slightly helps tree-based models, harms LSTM, and gives the largest improvement for
TFT. For TFT, adding full sentiment features reduces MAE from 0.1061 to 0.0573, showing that sentiment is useful
when the model can selectively use it.
Delivery cue: Say “selectively use it” clearly; this is the key interpretation.
Slide 14 - Explainability Target: 0:40
The SHAP analysis explains what drives the predictions. The strongest feature is demand from one week ago,
confirming the importance of recent momentum. The next important group is Google Trends search interest,
including Dubai Hotels, Dubai Tourism, and Visit Dubai. Sentiment features contribute, but they refine the forecast
rather than replace demand and search variables.
Delivery cue: Use this slide to show that the model is interpretable, not a black box.

Slide 15 - Transformer attention and forward forecasts Target: 0:35

The TFT model was also used for forward forecasts in April 2026. The predicted demand index is around 0.63 across
the 7-day, 14-day, and 30-day horizons, and all three are classified as peak season. This suggests a stable high-
demand period and shows how the model could support staffing, pricing, and capacity decisions.
Delivery cue: Connect the forecast to real managerial action.

Slide 16 - Discussion and practical implications Target: 0:45

The main practical lesson is that model choice should depend on the forecasting context. TFT is strongest when
long-range patterns and multiple behavioural signals matter. ARIMA remains strong when the task is short-term local
tracking. Therefore, tourism forecasting should not depend only on a headline score; it should use transparent
validation and choose models according to operational needs.
Delivery cue: Summarise the intellectual contribution in one clear message.

Slide 17 - Limitations and future work Target: 0:40

This study has three main limitations. First, the demand target is based on Google Trends rather than official arrivals
or hotel occupancy. Second, the review data is highly positive, which limits sentiment variation. Third, the study
focuses only on Dubai. Future work should use official tourism data, multiple destinations, additional external
variables such as events and flight bookings, and newer baselines such as iTransformer and N-BEATS.
Delivery cue: Sound balanced and confident; limitations strengthen the work.

Slide 18 - Conclusion Target: 0:45

To conclude, this dissertation shows that a Dubai-specific multisource forecasting framework can improve short-term
tourism demand prediction. TFT achieved the best held-out performance, with MAE of 0.0176 and R squared of
0.9518, while ARIMA performed best under rolling-origin validation. The main takeaway is that search behaviour and
BERT-based sentiment are valuable when combined with strong temporal modelling and transparent evaluation.
Thank you for your attention; I welcome your questions.
Delivery cue: End slowly and confidently. Do not add extra points after “questions”.

Slide 19 - References Target: 0:05

These are the key academic references used in the dissertation. I have included them for completeness and will be
happy to discuss them during questions.
Delivery cue: Do not read the references; move directly to Q&A.
Short Q&A Preparation Notes
Q: Why use Google Trends instead of official arrivals?
A: Because the study needed a public, reproducible, weekly proxy. Official arrivals or occupancy would be
stronger and are recommended for future work.
Q: Why did ARIMA beat TFT in rolling-origin validation?
A: Rolling-origin validation rewards local short-window tracking. ARIMA can exploit local stationarity, while TFT
needs more historical context to exploit multimodal patterns.
Q: What is the main value of sentiment?
A: Sentiment is not the main driver. It works as a complementary behavioural signal, especially when used by TFT
with attention and gating mechanisms.
Q: What is the strongest result?
A: TFT achieved the best held-out test performance, but the most defensible conclusion is that model choice
depends on validation protocol and operational scenario.