Trained Models, Not a Chat Wrapper
A general chatbot can talk about the market, but it was not trained to trade it. Tick Forward uses models trained on market data, so decisions are faster, cheaper, and repeatable.
AI / Finance / Local models
Tick Forward
A trading bot that trains its own models, reads the market, and makes the call.
The Numbers
0.925
AUC, Ridge Logistic Model
$5,000
Position Cap Per Trade
$250
Daily Loss Limit
Paper-trading results will be posted once the bot has a full track record.
The Problem
Markets move all day, and good trading asks for the same things every time: research, a plan, and the discipline to follow it. Doing that by hand is slow, emotional, and easy to skip when life gets busy. Most people either miss the setup or talk themselves into a bad one.
Currently, the problem is that Day Trading as a practice almost always loses in the long run, and that is just because of the volatility of the markets. The AI has not yet made money in any version of it or in any trial run not because it's bad or its training is bad, but rather because I haven't found a way to apply it to financial data in a way that puts it in a position to win. I'm currently trying to change the way the AI approaches the data, and change the way that Qwen and Laya interact with each other, and the market. The problem is I want to make money from the stock market with AI in a way that can outperform a computer script with pre-written rules, and have not been able to find that path yet. If you have any ideas, PLEASE contact me.
I made a report that walks you through what I've done so far in this project, and the many different angles I've come at it from.
Want to make reports like this with AI? Check out my skill.
Key Decisions
Multiple Types of Models
LLMs and SLMs are pretty slow for decisions on a market that is rapidly changing moment to moment. When every second matters, you need quick decisions. Laya is an open-source system 1 model (Like Jev) that can be run locally and is very fast. The catch is that it needs to be calibrated and sometimes trained so that it can make the right decisions confidently, which is where much of my learning is coming from.
Automation With Guardrails
Automation is only useful if it is safe. The model proposes and the rules dispose: risk limits and position sizing sit between the signal and the order, with a clear testing path before real money is involved.
Under the Hood
Numbers to a Logistic, Text to Laya
A ridge logistic model read the market snapshot numbers far better than Laya did (regime AUC 0.925 vs 0.631), so numbers go to the logistic. Laya reads text and breaks ties where the logistic is unsure. New kinds of data go to the logistic first.
Fine-Tuning Qwen With QLoRA
I fine-tuned Qwen3-4B-Instruct with 4-bit QLoRA in PyTorch (peft and bitsandbytes) to read a headline and call its direction. The answer is one word, scored from the first-token probabilities, so the base and tuned models compare without any sampling noise.
Honest Evaluation
The tuned model jumped from 71% to 88% accuracy on public sentiment data, but on 6,000 held-out Benzinga headlines labelled by what the stock did next, the AUC gain was +0.004 with an interval spanning zero. That is not a win, and the report says so. A model only counts if it beats the base on time-split data it never trained on.
Lookahead and Cost Guardrails
Headlines must be created before the entry open, and the tuned model is only tested after its last training day. A fitted Laya on 25 large caps reached ECE 0.02 to 0.06 but no measured edge. Hard limits cap each position at $5,000 and daily loss at $250, on a paper account until the evidence is there.
What's Next
The bot is still in development. Next up is a longer live track record, more strategies,
and tighter risk controls.
In the meantime, meet Lamp Light,
my personal document reader.
Tech Stack
Python
Trained ML models
PyTorch
QLoRA (peft, bitsandbytes)
Qwen3-4B
Laya
Ridge logistic baselines
Alpaca paper trading
Day-block bootstrap
Market data APIs
Automated order execution