Why I stopped trading and started tendering with AI

Description and use

The author moved from algorithmic trading to AI‑driven tendering in digital advertising, building a PyTorch model to predict impression value and bid accordingly.

Problem

Algorithmic trading is a zero‑sum game where profitable patterns are quickly arbitraged away, leading to diminishing edge and high effort‑to‑reward ratio.

Outcome

The AI‑driven tendering system replaces unpredictable trading with a value‑creating bidding strategy, yielding consistent profits and reducing the need for constant optimization.

Workflow

• 1. Collect bid request data from a DSP API (user demographics, device type, ad size, time, historical CTR/CVR).
• 2. Encode categorical features with an embedding layer.
• 3. Feed embeddings through a feedforward neural network (two hidden layers, ReLU activations).
• 4. Train the model with MSE loss on historical bid data, retraining daily for continual learning.
• 5. Apply exploration‑exploitation by adding a small random factor to predicted bids.
• 6. In the bidding engine, cap the predicted bid at the DSP’s max_bid and submit via the DSP API.

Evidence

I sunk countless hours (and a not‑insignificant amount of capital) into trying to predict the market.

Source

The market, it turns out, is remarkably good at not being predictable.

Source

I built an AI that can accurately predict the value of an impression to an advertiser and bid accordingly, without necessarily knowing the future price of the impression.

Source