Day-ahead probabilistic forecasts for the Turkish electricity market, with the reasons behind them, tested the way they would have been used.
Puant is Turkish for the evening peak hour, when the grid is under the most pressure. Every morning, before the day-ahead market closes at 12:30, Puant forecasts tomorrow's hourly market clearing price (PTF), consumption, wind and solar generation and the grid's carbon intensity, each as a 10–50–90 % range. It then explains why tomorrow's price is high or low, lets you rerun it under different weather, and simulates the auction on EPİAŞ's real supply and demand curves.
Open the live dashboard → · Türkçe özet
Contents: Results · What the dashboard shows · How it works · Limitations · Running it · FAQ · Türkçe özet · Citing
- Beats the official forecasts in an honest backtest: TEİAŞ's load plan by 25 % and EPİAŞ's wind forecast by 34 %. Price error is 28 % below the naive benchmark.
- Calibrated uncertainty: the 80 % bands hold 76–80 % of hours out of sample.
- Explains itself with exact Shapley values over driver groups (wind, demand, solar, gas, hydro, calendar, recent prices).
- Simulates the auction on EPİAŞ's published supply and demand curves, price cap included.
- Puts a value in lira on the forecast: a battery trading on it captures 82 % of the perfect-foresight profit.
- Runs itself: GitHub Actions forecasts every morning and grades each forecast once the day is over.
- No GPU, no paid data: public EPİAŞ and Open-Meteo data, scikit-learn, and a static page with no dependencies.
Rolling backtest from July 2024 to October 2026 (20,000+ hours). The models were retrained on the first of every month using only data before that day, and every day was forecast with the weather forecast issued the day before, never the weather that actually happened. The benchmark is the best official or simple forecast available at the same time.
| Forecast | Puant error (MAE) | Best benchmark | Benchmark error | Improvement | 80 % band held |
|---|---|---|---|---|---|
| Price | 311 TL/MWh | same hour today | 433 | 28 % | 79 % |
| Consumption | 893 MWh | TEİAŞ load plan | 1,185 | 25 % | 80 % |
| Wind | 656 MWh | EPİAŞ wind forecast | 1,000 | 34 % | 77 % |
| Solar (licensed) | 108 MWh | same hour last week | 200 | 46 % | 76 % |
| Carbon intensity | 22 g/kWh | same hour last week | 38 | 43 % | 79 % |
What it is worth. A 10 MWh / 5 MW battery (90 % round trip) that planned each day on Puant's median price forecast and was paid actual prices would have earned 15.4 million TL over 844 days: 82 % of what perfect foresight could have earned, and 2.0× a fixed "charge late morning, discharge in the evening" rule.
- Tomorrow's prices with an 80 % band, against today's prices.
- Why tomorrow costs what it does. The gap between tomorrow and a normal day (the last 28 days at the same hour) is split among wind, demand and temperature, sun, gas, hydro, calendar and recent prices with exact Shapley values over the groups. The parts add up to the gap, and no group takes credit for another's effect.
- What if. Tomorrow's price with temperature shifted ±4 °C (the consumption model moves demand with it), wind ±50 % and gas ±20 %.
- Market simulator. EPİAŞ publishes the aggregate supply and demand curves of every hourly auction. Their crossing reproduces the actual clearing price exactly in most hours (within a few percent otherwise: block bids are not in the curves). Shifting a curve answers questions such as "2,000 MW less wind" directly from real bids, price cap included.
- Cheapest and cleanest hours for flexible demand such as an EV, a washing machine or an industrial load.
- The battery plan for tomorrow and its backtested earnings.
- A scoreboard of the backtest and of every live forecast, graded once the day is over. Only forecasts made before the 12:30 market close count.
Information set. The forecast for day D+1 is made on the morning of D. Day D's prices are known (published on D−1); consumption and generation are known up to the morning of D, so their lags start at D−1. TEİAŞ's load plan and EPİAŞ's wind forecast for D+1 are published before the market closes. A test rewrites a forecast day's own prices, consumption and generation and checks that none of its features change.
Data. EPİAŞ Transparency Platform (prices, real-time
consumption and generation by source, load plan, wind forecast, gas reference price) and
Open-Meteo (ERA5 reanalysis, and archived day-ahead forecasts from their
previous-runs API). Temperature is weighted by the population of 26 provinces. Wind at hub height and
solar irradiance are weighted by the installed capacity of wind and solar farms mapped in
OpenStreetMap, clustered into 12 sites each
(scripts/build_sites.py). Wind speed is turned into a turbine power curve before averaging.
Models. One histogram gradient-boosting model per target and quantile (scikit-learn). Three choices make them hold up out of sample:
- Prices relative to their recent level. Lira prices rose roughly tenfold over 2019–2022, so the target and the price features are ratios to the mean price of the last 28 days.
- Output per unit of weather for wind and solar. Grid-connected solar nearly tripled since 2023 and trees cannot extrapolate. The models forecast output divided by the recent yield (MWh per unit of irradiance or wind power index over the last 2–4 weeks), which carries capacity growth forward automatically. This alone cut the solar error from 490 to 108 MWh.
- Official forecasts as inputs. The consumption model sees TEİAŞ's load plan and the wind model sees EPİAŞ's forecast, and both learn to correct them.
Calibrated bands. Quantile models trained on reanalysis weather are overconfident when fed a forecast. Training on archived day-ahead forecasts where they exist, plus split-conformal scaling on the last 90 days of out-of-sample errors, brings the 80 % bands to 76–80 % coverage.
Carbon. Hourly intensity of domestic generation from the fuel mix (IPCC fuel factors over typical plant efficiencies; Turkish geothermal counted at 0.38 t/MWh for its reservoir CO₂). The result, about 470 g/kWh on average, is in line with published estimates (Ember).
- The explanation is the model's view: which inputs moved its forecast. It is not a causal claim about the market.
- "Solar" is the real-time measured, licensed generation that EPİAŞ publishes. Rooftop and unlicensed plants appear only as lower consumption.
- Carbon intensity covers direct emissions of domestic generation. Imports and life-cycle emissions are excluded.
- The simulator uses the published aggregate curves. Block and flexible bids are not in them, and shifting a curve assumes everyone else bids the same.
- Not investment or trading advice.
pip install -e ".[model,dev]"
cp .env.example .env # add your EPİAŞ Transparency Platform login
python -m puant.data --weather # history since 2019 (weather since 2021), ~20 MB
python -m puant.backtest # the rolling backtest (~20 min on 10 cores)
python -m puant.live # tomorrow's forecast -> site/data/*.json
python -m http.server -d site # the dashboard
pytestThe Transparency Platform API needs a free EPİAŞ account. The credentials are read
from .env locally and from repository secrets in CI, and never enter git.
Automation. .github/workflows/daily.yml runs at 09:30 Turkish time. It restores the raw data
from the Actions cache, updates it, forecasts tomorrow, commits the forecast to data/forecasts/
for the live scoreboard, and deploys the dashboard to GitHub Pages. backtest.yml reruns the
backtest every month. Set EPIAS_USERNAME and EPIAS_PASSWORD as repository secrets.
src/puant/
epias.py EPİAŞ client (login ticket, chunking, throttling)
weather.py Open-Meteo: population- and capacity-weighted drivers
data.py incremental parquet store
calendar.py holidays, Bayram eves, bridge days
features.py as-of-morning features and targets
models.py quantile models, rolling backtest, conformal bands
explain.py Shapley attribution, scenario grid
market.py auction simulator on supply/demand curves
battery.py battery arbitrage (dynamic programming)
carbon.py carbon intensity, best hours
live.py tomorrow's forecast for the dashboard
site/ static dashboard (no dependencies)
Can I use this to trade? It is a research and transparency project, not trading advice. The backtest is honest, but past accuracy does not guarantee future accuracy, and real trading has costs, collateral and imbalance risk that this ignores.
Why not deep learning? With about 20,000 hourly test points and strong tabular features (calendar, official forecasts, weather, recent prices), gradient-boosted trees are accurate, fast to retrain every month on a CPU, and easy to explain. What made the difference was the information set, the capacity-free targets and calibration, not the model class.
Why are prices modelled as ratios? Lira prices rose about tenfold between 2019 and 2022. In absolute terms, a model trained on 2020 prices would be badly off in 2025; ratios to the recent mean travel across inflation regimes.
How do I forecast another market? Replace epias.py and data.py with a client for that
market's data, adjust calendar.py and the provinces in weather.py; the rest is
market-agnostic.
The daily job failed. What now? Check the EPİAŞ secrets first; their login tickets expire, and a changed password must be updated in the repository secrets. A cold cache downloads the full history, and Open-Meteo's free quota may cut the weather short; the next run resumes from where it stopped.
Puant, Türkiye elektrik piyasası için gün öncesi olasılıksal tahmin sistemidir. Her sabah gün öncesi piyasası (GÖP) 12:30'da kapanmadan önce, yarının saatlik piyasa takas fiyatı (PTF), elektrik tüketimi, rüzgâr ve güneş üretimi ile şebekenin karbon yoğunluğu için %10–50–90 aralıklı tahmin üretir ve canlı panoda yayımlar.
- Dürüst geriye dönük test (Temmuz 2024 – Ekim 2026): model her ay yalnızca geçmiş veriyle yeniden eğitildi ve her gün bir gün önce yapılmış hava tahminiyle tahmin edildi. Sonuç: PTF tahmininde naif kıyasa göre %28, tüketimde TEİAŞ yük tahmin planına göre %25, rüzgârda EPİAŞ rüzgâr tahminine göre %34 daha düşük hata.
- "Yarın elektrik neden pahalı?" Yarın ile normal bir gün arasındaki fiyat farkı rüzgâr, talep, güneş, doğalgaz, hidro, takvim ve son günlerin fiyatı arasında Shapley değerleriyle bölüştürülür.
- Senaryo: sıcaklık, rüzgâr ve doğalgaz fiyatı değişirse yarının fiyatı ne olur?
- Piyasa simülatörü: EPİAŞ'ın yayımladığı gerçek arz-talep eğrileriyle "2.000 MW daha az rüzgâr fiyatı ne kadar artırır?" sorusunun yanıtı.
- En ucuz ve en temiz saatler: elektrikli araç şarjı, beyaz eşya ya da sanayi yükü için.
- Batarya arbitrajı: 10 MWh'lik bir batarya bu tahminle kusursuz öngörü kazancının %82'sini yakalardı (844 günde 15,4 milyon TL).
- Veriler: EPİAŞ Şeffaflık Platformu, Open-Meteo, OpenStreetMap. Kod MIT lisanslıdır.
Anahtar kelimeler: elektrik fiyat tahmini, PTF tahmini, gün öncesi piyasası, EPİAŞ şeffaflık platformu, elektrik tüketim tahmini, rüzgâr ve güneş üretim tahmini, karbon yoğunluğu, enerji piyasası makine öğrenmesi, batarya arbitrajı.
If Puant is useful in your work, please cite it:
@software{tunali_puant_2026,
author = {Tunali, Emir},
title = {Puant: day-ahead probabilistic forecasts for the Turkish electricity market},
year = {2026},
url = {https://github.com/emrtunali/puant}
}Issues and pull requests are welcome: bug reports, new benchmarks, other markets.
Code: MIT. EPİAŞ Transparency Platform data is public. Weather data © Open-Meteo, CC BY 4.0 (ERA5: Copernicus Climate Change Service). Plant locations © OpenStreetMap contributors, ODbL.