EPFL Quantum Hackathon 2026 • Quandela Challenge
Teaching photons to predict the market so we can finally sleep.
A swaption volatility surface is a 2D grid indexed by tenor and maturity. Providing 224 high-dimensional features per day, the challenge demands two primary tasks:
-
Predictive Forecasting — Forecasting the complex non-linear surface shifts for the next
$H$ days smoothly. -
Data Imputation Reconstruction — Real markets suffer from outages, resulting in
NaNentries across the option grid. The model must analytically infer missing surface areas by relying on deeply correlated temporal topologies.
The Quantum Solution: We compress the manifold via PCA and process the temporal dynamics using a entirely newly adapted Hybrid Photonic Temporal Quantum Reservoir Computer (HPT-QRC) architecture. Inspired by Li et al. (2024), we took their qubit-based framework and adapted it to a purely Photonic Quantum Reservoir powered by MerLin by Quandela.
You can check out our interactive showcase website here: https://qedi-qpfl.vercel.app, featuring all challenge details and interactive visualization of our solution.
Instead of a standard QRC, our best solution is the state-of-the-art HPT-QRC pipeline (implemented as our main architecture in temporal_qrc.py):
-
Preprocessing: Raw 224D Market Surface
$\rightarrow$ StandardScaler+PCA (5D)$\rightarrow$ Rolling 5-Day Window (1×25) - Dedicated Memory Modes: Instead of mapping data to all spatial modes simultaneously, our temporal array uses 5 input modes (phase encoded) and 3 dedicated memory modes (unencoded loop). The memory modes continuously accumulate historical state contexts across 5 time steps through serial phase mixing.
- Virtual Nodes: We sample the evolving physical system at multiple structural post-processing depths (Depths 1, 2, 3). These Virtual Nodes emulate capturing chronological measurement sub-intervals, massively expanding our temporal feature dimensionality without adding physical photon bounds.
- Ensemble LexGrouping Compression: Instead of measuring impossibly vast raw Fock states, we group the probability vectors via LexGrouping across 3 random seeds × 3 virtual depths (= 9 Circuits).
-
Direct Target Ridge Forecaster: By augmenting 90 Quantum Features with the 25 Classic Features, extracting the most prominent Non-linear Mutual Information quantum channels, an
$L2$ regularized Ridge Regression projection ($\alpha=10.0$ ) predicts the consecutive future states.
Requires torch, pcvl, and standard ML arrays (numpy, pandas, scikit-learn).
# First create the environment and install dependencies
conda create -n quandela python=3.10
conda activate quandela
pip install -r requirements.txtOur best model, the Hybrid Photonic Temporal QRC, along with a naive algorithm (repeat last known day) used as a baseline, is localized in temporal_qrc.py. Run this to reproduce our winning results:
python temporal_qrc.pyWe built multiple robust models to benchmark against our champion HPT-QRC architecture:
final_model.py: Runs our standard Photonic Linear QRC without dedicated memory modes.train_final.py: Runs classical and hybrid benchmarks, including Classical LSTM, Classical Random Forest, QSVR, and Hybrid QNN.
# Evaluate the base Photonic Linear QRC
python final_model.py
# Evaluate classical/hybrid benchmarking models
python train_final.pyBy executing the novel Hybrid Photonic Temporal QRC pipeline, we successfully beat our underlying standard QRC framework and left classical baselines absolutely obsolete.
| Model | RMSE Error |
|---|---|
| QSVR | 0.0233 |
| Hybrid QNN | 0.0083 |
| LSTM | 0.0073 |
| Photonic Linear QRC | 0.0065 |
| Hybrid Photonic Linear QRC | 0.0028 |
| 🥇 Hybrid Photonic Temporal QRC (HPT-QRC) | 0.0021 |
Maintains incredibly reliable sub-10% projection accuracy out to a 6-Day cascading prediction envelope.
Proudly built during the 24-hour EPFL Quantum Hackathon 2026.
|
|
| Middle East Technical University | Bilkent University |