Datasets:

rathishTharusha
/

srilanka-tourist-arrivals

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 3 new columns ({'Colombo_Temp_C', 'Date', 'Colombo_Precip_mm'}) and 4 missing columns ({'Year', 'Day', 'Count', 'Month'}).

This happened while the csv dataset builder was generating data using

hf://datasets/rathishTharusha/srilanka-tourist-arrivals/Colombo_Weather_Daily_2023_2025.csv (at revision 33f7854600b4122ac6c3758c3fc5a185ae1ee85f), ['hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/0_SLTDA_Arrivals_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Colombo_Weather_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Google_Trends_Localized_Weekly_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Google_Trends_Normalized_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/USD_LKR_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Yandex_Flights_Colombo_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Yandex_Tour_SriLanka_2023_2025.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              Date: string
              Colombo_Temp_C: double
              Colombo_Precip_mm: double
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 652
              to
              {'Year': Value('int64'), 'Month': Value('int64'), 'Day': Value('int64'), 'Count': Value('string')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 3 new columns ({'Colombo_Temp_C', 'Date', 'Colombo_Precip_mm'}) and 4 missing columns ({'Year', 'Day', 'Count', 'Month'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/rathishTharusha/srilanka-tourist-arrivals/Colombo_Weather_Daily_2023_2025.csv (at revision 33f7854600b4122ac6c3758c3fc5a185ae1ee85f), ['hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/0_SLTDA_Arrivals_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Colombo_Weather_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Google_Trends_Localized_Weekly_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Google_Trends_Normalized_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/USD_LKR_Daily_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Yandex_Flights_Colombo_2023_2025.csv', 'hf://datasets/rathishTharusha/srilanka-tourist-arrivals@33f7854600b4122ac6c3758c3fc5a185ae1ee85f/Yandex_Tour_SriLanka_2023_2025.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Year int64	Month int64	Day int64	Count string
2,023	1	1	2,246
2,023	1	2	3,633
2,023	1	3	2,982
2,023	1	4	2,694
2,023	1	5	3,512
2,023	1	6	3,035
2,023	1	7	2,773
2,023	1	8	2,912
2,023	1	9	3,806
2,023	1	10	3,239
2,023	1	11	2,702
2,023	1	12	4,378
2,023	1	13	3,321
2,023	1	14	2,937
2,023	1	15	3,183
2,023	1	16	4,009
2,023	1	17	3,253
2,023	1	18	3,746
2,023	1	19	3,973
2,023	1	20	3,506
2,023	1	21	3,126
2,023	1	22	3,036
2,023	1	23	3,995
2,023	1	24	3,146
2,023	1	25	2,973
2,023	1	26	3,879
2,023	1	27	3,451
2,023	1	28	2,880
2,023	1	29	3,130
2,023	1	30	3,915
2,023	1	31	3,174
2,023	2	1	3,121
2,023	2	2	3,940
2,023	2	3	4,073
2,023	2	4	3,313
2,023	2	5	3,594
2,023	2	6	4,638
2,023	2	7	3,827
2,023	2	8	3,688
2,023	2	9	4,591
2,023	2	10	4,722
2,023	2	11	3,684
2,023	2	12	3,546
2,023	2	13	4,510
2,023	2	14	3,438
2,023	2	15	3,359
2,023	2	16	4,377
2,023	2	17	4,334
2,023	2	18	3,407
2,023	2	19	3,559
2,023	2	20	4,619
2,023	2	21	3,231
2,023	2	22	3,542
2,023	2	23	4,630
2,023	2	24	3,988
2,023	2	25	3,732
2,023	2	26	3,073
2,023	2	27	4,041
2,023	2	28	3,062
2,023	3	1	4,035
2,023	3	2	4,015
2,023	3	3	4,806
2,023	3	4	3,687
2,023	3	5	3,347
2,023	3	6	4,473
2,023	3	7	3,210
2,023	3	8	3,513
2,023	3	9	3,956
2,023	3	10	4,338
2,023	3	11	5,846
2,023	3	12	3,203
2,023	3	13	5,409
2,023	3	14	3,199
2,023	3	15	3,845
2,023	3	16	4,404
2,023	3	17	3,764
2,023	3	18	3,575
2,023	3	19	3,620
2,023	3	20	5,081
2,023	3	21	3,257
2,023	3	22	3,703
2,023	3	23	3,780
2,023	3	24	4,713
2,023	3	25	4,564
2,023	3	26	4,369
2,023	3	27	4,353
2,023	3	28	2,925
2,023	3	29	3,627
2,023	3	30	4,274
2,023	3	31	4,604
2,023	4	1	4,238
2,023	4	2	3,695
2,023	4	3	4,548
2,023	4	4	3,260
2,023	4	5	3,211
2,023	4	6	4,071
2,023	4	7	3,888
2,023	4	8	3,158
2,023	4	9	2,425
2,023	4	10	3,550

End of preview.

Sri Lanka Daily Tourist Arrivals & Search Intent Dataset (2023–2025)

Dataset Description

This dataset provides a unified, daily time-series panel designed for forecasting international tourist arrivals to Sri Lanka over the post-crisis recovery period of January 2023 to December 2025. It integrates five disparate data streams to capture the macroeconomic, environmental, and behavioural signals that drive tourism demand.

The dataset was curated as part of the MERCon 2026 research paper that benchmarks seven forecasting architectures (SARIMA, SARIMAX, XGBoost, LSTM, SVR, Grid Search SVR, and a Sequential Hybrid) and explains their predictions using cross-model SHAP analysis.

Key Characteristics

Coverage: January 1, 2023 – December 31, 2025 (1,096 daily observations)
Target variable: Daily international tourist arrivals (SLTDA official counts)
Exogenous signals: Exchange rate, weather, Google Trends (4 source markets), Yandex Wordstat (Russian market)
Source markets covered: Global (EN), Russia (RU), United Kingdom (GB), Germany (DE), India (IN)
Train/test split boundary used in paper: 2024-06-29

Paper Citation

If you use this dataset in your research, please cite:

@inproceedings{perera2026searchintent,
  title     = {The Effect of Search Intent Data on Predicting Post-Crisis Daily Tourist Arrivals in Sri Lanka},
  author    = {Perera, Tharusha and Mahanama, Janith and Kumarasinghe, Maleesha and
               Udurawana, Bimsara and Nawarathna, Praveen and Narasinghe, Patalee and
               Athukorala, Nathali and Wickramanayake, Sandareka and {de Silva}, Nisansa},
  booktitle = {Proceedings of the Moratuwa Engineering Research Conference (MERCon)},
  year      = {2026},
  institution = {University of Moratuwa, Sri Lanka}
}

Code repository: github.com/rathishTharusha/srilanka-tourist-arrivals-prediction

Sub-Dataset Files & Data Dictionary

The dataset consists of 7 CSV files. Each file is documented below with a full column-level data dictionary.

File 1: `SLTDA_Arrivals_Daily_2023_2025.csv`

Description: The primary target variable. Official daily counts of international tourist arrivals to Sri Lanka, sourced from the Sri Lanka Tourism Development Authority (SLTDA).

Source: Sri Lanka Tourism Development Authority (SLTDA) Annual Statistical Reports
Temporal coverage: 2023-01-01 to 2025-12-31
Granularity: Daily
Shape: 1,096 rows × 4 columns

Loading note: The Count column is stored with comma-thousands separators (e.g., "2,246"). Parse as: pd.to_numeric(df['Count'].str.replace(',', ''), errors='coerce').

Column	Data Type	Unit	Description	Value Range	Missing Values
`Year`	int	—	Calendar year of the observation	2023–2025	None
`Month`	int	—	Calendar month (1 = January, 12 = December)	1–12	None
`Day`	int	—	Calendar day of the month	1–31	None
`Count`	string (comma-formatted)	arrivals / day	Raw total international tourist arrivals for that calendar day. Stored as a comma-separated string from the original SLTDA export. Must be parsed before use.	~2,000–12,000	None

Derived feature used in paper: $y_t = \log_{10}(\text{Count}_t + 10)$ This log₁₀ transformation (offset +10 to handle near-zero values) stabilises variance and compresses the operating range to approximately 3.3–4.1.

File 2: `USD_LKR_Daily_2023_2025.csv`

Description: Daily USD/LKR (US Dollar to Sri Lankan Rupee) exchange rate. Captures the macroeconomic cost-of-travel effect from the perspective of foreign visitors.

Source: Yahoo Finance (Ticker: USDLKR=X)
Temporal coverage: 2023-01-02 to 2025-12-31 (weekdays only)
Granularity: Daily (business days only; no weekend entries)
Shape: ~780 rows × 2 columns (after skipping Yahoo Finance metadata rows)

⚠️ Loading Note: This file has 3 metadata rows at the top inherited from Yahoo Finance's CSV export format (rows 0–2: Price/USD_LKR_Rate, Ticker/USDLKR=X, Date/<blank>). Read with:
df = pd.read_csv('USD_LKR_Daily_2023_2025.csv', skiprows=3, names=['Date', 'USD_LKR_Rate'])

Column	Data Type	Unit	Description	Value Range	Missing Values
`Date`	string (YYYY-MM-DD)	—	Trading date. Weekends and Sri Lankan bank holidays are absent.	2023-01-02 to 2025-12-31	None (but weekend gaps exist)
`USD_LKR_Rate`	float64	LKR per 1 USD	Closing exchange rate for that trading day. Higher values = weaker rupee (more expensive for foreign tourists to convert).	~295–375	None in trading days; forward-fill applied for weekend/holiday gaps in the merged panel

Context: The LKR depreciated sharply during the 2022 economic crisis and partially stabilised during the 2023–2025 recovery period covered by this dataset.

File 3: `Colombo_Weather_Daily_2023_2025.csv`

Description: Daily meteorological observations for Colombo, Sri Lanka. Weather conditions at the primary international arrival hub (Bandaranaike International Airport, Colombo) are used as a contextual control variable.

Source: Open-Meteo API (open-source weather data, ERA5 reanalysis)
Temporal coverage: 2023-01-01 to 2025-12-31
Granularity: Daily
Shape: 1,096 rows × 3 columns
Encoding: UTF-8

Column	Data Type	Unit	Description	Value Range	Missing Values
`Date`	string (YYYY-MM-DD)	—	Calendar date	2023-01-01 to 2025-12-31	None
`Colombo_Temp_C`	float64	°C	Daily mean air temperature at Colombo	~22–32 °C	None
`Colombo_Precip_mm`	float64	mm	Total daily precipitation at Colombo	0–200+ mm	None

SHAP finding from paper: Both temperature and precipitation contributed negligible SHAP values across all models, confirming they are statistical noise at daily granularity for arrival prediction. They are retained for completeness and for future studies with finer spatial resolution.

File 4: `Google_Trends_Normalized_Daily_2023_2025.csv`

Description: Localised Google Trends Search Volume Indices (SVIs) for tourism-related keywords, segmented by source market. Contains signals for four geographic markets: Global (English), United Kingdom, Germany, and India.

Source: Google Trends API (via pytrends or manual download)
Temporal coverage: 2023-01-01 to 2025-12-31
Granularity: Daily (normalised from weekly raw data via interpolation)
Shape: 1,096 rows × 12 columns
Encoding: UTF-8 (Cyrillic characters in Russian columns — see note below)

⚠️ Encoding Note: This file contains Cyrillic characters in the Russian-market column names (RU_*). Read with encoding='utf-8'. The Russian Google Trends columns (RU_SriLankatur, RU_aviabiletKolombo) are zero-variance across the entire dataset — they were found to contain no signal. The Yandex dataset files should be used for the Russian market instead.

⚠️ Normalisation Note: Google Trends reports normalised SVIs (0–100) where 100 = peak search interest within the selected time frame. Values are relative, not absolute query counts.

Column	Data Type	Unit	Locale	Keyword (English)	Value Range	Notes
`date`	string (YYYY-MM-DD)	—	—	—	2023-01-01 to 2025-12-31	Index column
`_SriLankatravel`	float64	SVI (0–100)	Global (EN)	"Sri Lanka travel"	0–100	Global English search interest
`_SriLankaflights`	float64	SVI (0–100)	Global (EN)	"Sri Lanka flights"	0–100	Global English search interest
`_SriLankaVisa`	float64	SVI (0–100)	Global (EN)	"Sri Lanka Visa"	0–100	Global English search interest
`RU_SriLankatur`	float64	SVI (0–100)	Russia (RU)	"Шри-Ланка тур" (Sri Lanka tour)	0 (constant)	⚠️ Zero-variance — no predictive signal. Use `Yandex_Tour_SriLanka_2023_2025.csv` instead
`RU_aviabiletKolombo`	float64	SVI (0–100)	Russia (RU)	"авиабилеты Коломбо" (flights Colombo)	0 (constant)	⚠️ Zero-variance — no predictive signal. Use `Yandex_Flights_Colombo_2023_2025.csv` instead
`GB_SriLankaholidays`	float64	SVI (0–100)	United Kingdom (GB)	"Sri Lanka holidays"	0–100	UK English search interest
`GB_SriLankaETA`	float64	SVI (0–100)	United Kingdom (GB)	"Sri Lanka ETA"	0–100	UK visa/electronic travel authority queries; nominal Granger evidence (raw p=0.0015)
`DE_SriLankaUrlaub`	float64	SVI (0–100)	Germany (DE)	"Sri Lanka Urlaub" (holiday)	0–100	German-language travel search
`DE_FlugeColombo`	float64	SVI (0–100)	Germany (DE)	"Flüge Colombo" (flights Colombo)	0–100	German-language flights search
`IN_SriLankatourpackage`	float64	SVI (0–100)	India (IN)	"Sri Lanka tour package"	0–100	Retained as exogenous feature (CCF r=−0.110, lag 31 days)
`IN_Colomboflight`	float64	SVI (0–100)	India (IN)	"Colombo flight"	0–100	Indian market flight search

File 5: `Google_Trends_Localized_Weekly_2023_2025.csv`

Description: Identical to File 4 but at weekly granularity (ISO week start dates). This file is the original raw weekly export from Google Trends before daily interpolation. It serves as a fallback source when the daily-interpolated version is unavailable.

Source: Google Trends API
Temporal coverage: 2023-01-01 to 2025-12-31
Granularity: Weekly (ISO week start, Monday-based)
Shape: 157 rows × 12 columns
Encoding: UTF-8

Column	Description
`date`	ISO week start date (YYYY-MM-DD, Mondays)
All other columns	Same 11 SVI columns as File 4, at weekly resolution

Usage in pipeline: In the notebook, the daily file (File 4) is preferred. If not found, the weekly file is loaded and forward-fill + backward-fill imputation is applied to produce daily values.

File 6: `Yandex_Tour_SriLanka_2023_2025.csv`

Description: Weekly absolute query volume from Yandex Wordstat for Russian-language tour-intent searches ("Шри-Ланка тур" — "Sri Lanka tour"). This is the strongest exogenous predictor identified in the study, achieving a cross-correlation of r = 0.423 with arrivals at a 27-day lag — representing the Russian market's booking horizon.

Original filename: Шри-Ланка тур 2023-2025.csv (renamed to ASCII for compatibility)
Source: Yandex Wordstat (yandex.com/wordstat) — Russian search engine
Temporal coverage: 2022-12-26 to 2026-01-04 (includes a 6-day pre-period and post-period)
Granularity: Weekly (ISO week start, Monday-based)
Shape: 158 rows × 4 columns
Encoding: UTF-8 with BOM (utf-8-sig)
Delimiter: Semicolon (;)

⚠️ Loading Note:

df = pd.read_csv('Yandex_Tour_SriLanka_2023_2025.csv',
                 sep=';', skiprows=1, header=None,
                 usecols=[0, 1], names=['Date', 'Yandex_RU_Tour'])
df['Date'] = pd.to_datetime(df['Date'], format='%d.%m.%Y', errors='coerce')
df['Yandex_RU_Tour'] = pd.to_numeric(
    df['Yandex_RU_Tour'].astype(str)
    .str.replace('\u00a0', '').str.replace(' ', ''), errors='coerce')

The Number of queries column uses non-breaking spaces (Unicode \u00a0) as thousands separators (e.g., "158 978" = 158,978).

Column	Data Type	Unit	Description	Value Range	Notes
`Week from`	string (dd.mm.yyyy)	—	ISO week start date in Yandex's European date format	26.12.2022 to 04.01.2026	Parse with `format='%d.%m.%Y'`
`Number of queries`	int	absolute query count	Total weekly Yandex searches for "Шри-Ланка тур" across all Russian devices	~100,000–700,000	Contains `\u00a0` (NBSP) thousands separators — must be stripped before parsing
`Percentage of total queries, %`	string	% of all Yandex queries	Share of this keyword relative to all Yandex queries that week. Decimal comma (`,`) used — convert with `.replace(',', '.')`	~0.003%–0.010%	Low absolute share but highly correlated with arrivals
(Column 4)	—	—	Metadata column from Yandex export (always NaN) — safe to ignore	—	Drop this column

Paper finding: Yandex_RU_Tour was retained as a primary exogenous feature in all ML models and SARIMAX. In the Hybrid XGBoost corrector, it achieved a normalised SHAP importance of 1.00, meaning it is the single most important feature for correcting SARIMAX's systematic errors during Russian demand spikes.

File 7: `Yandex_Flights_Colombo_2023_2025.csv`

Description: Weekly absolute query volume from Yandex Wordstat for Russian-language flight-booking searches ("авиабилеты Коломбо" — "flights to Colombo"). Captures transactional booking intent complementary to the tour-search signal.

Original filename: авиабилеты Коломбо 2023-2025.csv (renamed to ASCII for compatibility)
Source: Yandex Wordstat — Russian search engine
Temporal coverage: 2022-12-26 to 2026-01-04
Granularity: Weekly
Shape: 158 rows × 4 columns
Encoding: UTF-8 with BOM (utf-8-sig)
Delimiter: Semicolon (;)

⚠️ Loading Note: Same format as File 6. Use identical loading code, substituting value_name='Yandex_RU_Flights'.

Column	Data Type	Unit	Description	Value Range	Notes
`Week from`	string (dd.mm.yyyy)	—	ISO week start date	26.12.2022 to 04.01.2026	Parse with `format='%d.%m.%Y'`
`Number of queries`	int	absolute query count	Total weekly Yandex searches for "авиабилеты Коломбо"	~500–5,000	Contains `\u00a0` thousands separators
`Percentage of total queries, %`	string	% of all Yandex queries	Share relative to all Yandex queries	~0.00002%–0.00009%	Convert decimal comma before parsing
(Column 4)	—	—	Metadata column from Yandex export (always NaN)	—	Drop

Note: Yandex_RU_Flights is highly correlated with Yandex_RU_Tour (Pearson r = 0.92). Due to this multicollinearity, only Yandex_RU_Tour is used in the SARIMAX (V3) feature matrix, while both are included in the V1 matrix for tree-based and kernel models.

Preprocessing & Pipeline Notes

The following preprocessing steps are applied in the main notebook (tourist_arrivals_prediction_pipeline.ipynb) to produce the merged daily panel:

Date alignment: All sources are left-joined onto the SLTDA daily date index (1,096 rows, 2023-01-01 to 2025-12-31).
Forward-fill + Backward-fill imputation: Applied to all non-daily sources (USD/LKR weekends, weekly Google Trends and Yandex data) to produce a complete daily panel. No future information is introduced — bfill is applied only to handle the series start, not for future-filling.
Target transformation: Arrivals_Log10 = log10(Count + 10) applied to stabilise variance.
Leakage-safe scaling: MinMax scaling for visualisation is fitted exclusively on the training window (≤ 2024-06-29) and then applied to the full series. Model-level scaling (Robust Scaler for SVR) is performed inside the rolling window at each step.
Train/test split boundary: 2024-06-29 (training) / 2024-06-30 (test onset). All feature selection (Granger causality, CCF analysis) and hyperparameter tuning is strictly confined to training data.

Usage Example

import pandas as pd
import numpy as np

# --- 1. SLTDA Target ---
sltda = pd.read_csv('SLTDA_Arrivals_Daily_2023_2025.csv')
sltda['Date'] = pd.to_datetime(sltda[['Year', 'Month', 'Day']])
sltda['Count'] = pd.to_numeric(
    sltda['Count'].astype(str).str.replace(',', ''), errors='coerce')
sltda['Arrivals_Log10'] = np.log10(sltda['Count'] + 10)
sltda = sltda[['Date', 'Count', 'Arrivals_Log10']].set_index('Date')

# --- 2. USD/LKR ---
usd = pd.read_csv('USD_LKR_Daily_2023_2025.csv',
                  skiprows=3, names=['Date', 'USD_LKR_Rate'])
usd['Date'] = pd.to_datetime(usd['Date'], errors='coerce')
usd = usd.dropna(subset=['Date']).set_index('Date')
usd['USD_LKR_Rate'] = pd.to_numeric(usd['USD_LKR_Rate'], errors='coerce')

# --- 3. Weather ---
weather = pd.read_csv('Colombo_Weather_Daily_2023_2025.csv',
                      parse_dates=['Date'], index_col='Date')

# --- 4. Google Trends ---
trends = pd.read_csv('Google_Trends_Normalized_Daily_2023_2025.csv',
                     encoding='utf-8')
trends['Date'] = pd.to_datetime(trends['date'])
trends = trends.drop(columns=['date']).set_index('Date')

# --- 5 & 6. Yandex ---
def load_yandex(file_path, value_name):
    raw = pd.read_csv(file_path, sep=';', skiprows=1, header=None,
                      usecols=[0, 1], names=['Date', value_name])
    raw[value_name] = pd.to_numeric(
        raw[value_name].astype(str)
        .str.replace('\u00a0', '').str.replace(' ', ''),
        errors='coerce')
    raw['Date'] = pd.to_datetime(raw['Date'], format='%d.%m.%Y', errors='coerce')
    return raw.dropna().set_index('Date')

yandex_tour    = load_yandex('Yandex_Tour_SriLanka_2023_2025.csv',    'Yandex_RU_Tour')
yandex_flights = load_yandex('Yandex_Flights_Colombo_2023_2025.csv',  'Yandex_RU_Flights')

# --- 7. Merge & forward-fill ---
df = (sltda
      .join(usd, how='left')
      .join(weather, how='left')
      .join(trends, how='left')
      .join(yandex_tour, how='left')
      .join(yandex_flights, how='left'))

fill_cols = ['USD_LKR_Rate', 'Colombo_Temp_C', 'Colombo_Precip_mm',
             'Yandex_RU_Tour', 'Yandex_RU_Flights'] + list(trends.columns)
df[[c for c in fill_cols if c in df.columns]] = (
    df[[c for c in fill_cols if c in df.columns]].ffill().bfill())

print(f"Panel shape: {df.shape}")
print(df.head())

Structural Break Summary

The paper identified 4 structural breakpoints in the Sri Lankan daily arrivals series, creating 5 demand regimes:

Regime	Period	Mean log₁₀ Arrivals	Interpretation
1	Jan 2023 – Jul 2023	3.530	Early post-crisis recovery (fragile)
2	Jul 2023 – Nov 2023	3.664	Accelerating recovery (+4.1% at break)
3	Nov 2023 – Mar 2024	3.821	Peak recovery momentum (+3.8% at break)
4	Mar 2024 – Apr 2025	3.747	Established growth, slight softening (−3.5%)
5	Apr 2025 – Dec 2025	3.760	Novel demand plateau (−4.1%), out-of-training-distribution

Known Limitations & Caveats

Issue	Affected Files	Impact	Mitigation
Weekday-only exchange rate	`USD_LKR_Daily_2023_2025.csv`	~28% of days are weekend/holiday gaps	Forward-fill applied in pipeline
Weekly Yandex resolution	`Yandex_Tour_`, `Yandex_Flights_`	7-day step quantisation; all 7 days in a week share the same value after forward-fill	Acceptable for daily modelling; noted as a data granularity limitation in the paper
Russian Google Trends = zero variance	`Google_Trends_Normalized_Daily_2023_2025.csv` (columns `RU_*`)	These columns carry zero predictive information	Use Yandex files for Russian market signals
Non-breaking space in Yandex query counts	`Yandex_*.csv`	Raw integer field contains `\u00a0` thousands separators (e.g., `"158 978"`)	Strip NBSP characters before numeric conversion (see loading example)
Yahoo Finance metadata header rows	`USD_LKR_Daily_2023_2025.csv`	First 3 rows are Yahoo Finance export metadata, not data	Use `skiprows=3` when loading
Cyrillic column names	`Google_Trends_Normalized_Daily_2023_2025.csv`	Column name encoding may appear garbled in some editors	Read with `encoding='utf-8'`; rename columns as needed
Short post-crisis window	All files	Only ~550 days of training data available	Limits deep learning model capacity; see paper's limitations section

Ethical Considerations

All data used in this study is sourced from publicly available APIs and official government reports.
No personally identifiable information (PII) is present in any file. All observations are aggregate-level (daily totals or weekly indices).
The Yandex and Google Trends data represent aggregate, anonymised search interest indices, not individual user data.
The SLTDA arrival data is official government statistics published in public annual reports.

Dataset Authors

Name	Affiliation
Tharusha Perera	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Janith Mahanama	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Maleesha Kumarasinghe	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Bimsara Udurawana	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Praveen Nawarathna	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Patalee Narasinghe	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Nathali Athukorala	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Sandareka Wickramanayake	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka
Nisansa de Silva	Dept. of Computer Science & Engineering, University of Moratuwa, Sri Lanka

Downloads last month: 18

Total file size:

174 kB