Market Impact of Rocket Alerts on TASE (High-Frequency Study)

Minute-by-minute event study on the Tel Aviv Stock Exchange (TASE).
Raw order-book messages are reconstructed into executed trades and aggregated to 1-minute OHLCV/turnover; events are aligned to rocket alerts and estimated via a two-way fixed-effects panel.

Problem

Rocket alerts can affect prices and liquidity within minutes, yet public data rarely arrives as clean, aligned, high-frequency series. The goals are:

1. convert noisy order-book messages into uniform 1-minute security-level series;
2. estimate short-run abnormal returns and turnover around alerts while controlling for stock and time effects.

Methodology

1. Order book → trades
   Parse add/cancel/execute messages, reconstruct prints, reconcile volumes, enforce tick size, de-duplicate, validate timestamps/sequence.

2. Continuous 1-minute bars
   Aggregate to 1m OHLCV and turnover (₪), handle missing minutes, and mark open/close and halts.

3. Focus on market movers
   Use the most influential TASE-35 constituents (by index weight/turnover) to maximize signal-to-noise.

4. Event alignment
   Merge Home Front Command alerts and build a symmetric ±15-minute window per alert; exclude open/close and overlapping windows.

5. Controls for time patterns
   Include minute-of-day seasonality and event-time indicators (τ = −15..+15).

6. Panel models

   Returns (two-way FE):

   $$ r_{i,t} = \sum_{k=-15}^{+15} \gamma_k, D_{k,t} +\ \alpha_i +\ \delta_d +\ \theta_h +\ \varepsilon_{i,t} $$

   Turnover (log, HAC):

   $$ \log!\bigl(1+\mathrm{Turnover}^{\mathrm{TA35}}_{t}\bigr) = \sum_{k=-15}^{+15} \beta_k, D_{k,t} +\ \delta_d +\ \theta_h +\ \varepsilon_{t} $$

7. Quality control
   Sequence/time checks, volume reconciliation, tick-size enforcement, de-duplication, sparse-interval handling, outlier flags, validation of minute aggregates.

Results (high level)

• Returns: small but statistically significant negative move around the alert; trough ≈ −0.035% several minutes post-alert.
• Turnover: sharper immediate drop, ~6% per minute on average in the first five minutes; ≈ ₪1.45M cumulative over 15 minutes for the analyzed names.
• No immediate full rebound within the ±15-minute window.

Figures, robustness checks, and grouped-Wald tests appear in the notebooks and poster.

Repository layout

repo-root/
├─ notebooks/
│  ├─ Return_Panel_Model.ipynb
│  └─ Turnover_OLS-HAC_model.ipynb
├─ poster/
│  └─ Research_Poster_Hagai_BY.pdf
├─ src/
│  └─ config.py
├─ data/
│  ├─ README.md
│  └─ (raw/ and sample/ are created locally; git-ignored)
├─ download_data.py
├─ make_sample.py
├─ requirements.txt
└─ README.md

Quick start

## Quick start

git clone https://github.com/Hagai-BY/TASE-Red-Alert-Impact.git
cd TASE-Red-Alert-Impact

python -m venv .venv
# Windows:
. .venv\Scripts\Activate.ps1
# macOS/Linux:
# source .venv/bin/activate

pip install -r requirements.txt
jupyter lab  # or: jupyter notebook

Open notebooks/Return_Panel_Model.ipynb or notebooks/Turnover_OLS-HAC_model.ipynb and run all cells. If required CSVs are not found locally, the notebooks download them from the shared Google Drive folder into data/raw/.