How FundSpec's AI trading models are built and scored

FundSpec, 8 September 2026

Most products that promise AI trading signals describe the model as little as possible and its results even less. FundSpec takes the opposite view: a model that cannot show its work is not worth following. This guide explains what the models are, how they learn, how they are tested, and exactly what the record next to each one contains.

One model per ticker

FundSpec's trading models are reinforcement learning agents, and each one is trained on a single ticker. That is a deliberate choice. Stocks differ in volatility, liquidity, how they respond to the market and how they behave around their own events, and a model that has seen only one stock's history learns that stock's habits rather than an average of the market's.

An agent is a small neural network that looks at a window of recent price bars, daily bars for swing trading models and intraday bars for day trading models, and decides at each step whether to be long, short or flat. It has no opinion about the company. It has only the price path and the reward it earned the last time it acted.

Learning by trading a simulated market

Reinforcement learning trains the agent by letting it act. The training environment replays the ticker's history bar by bar. At each bar the agent calls the ticker bullish, bearish or neutral, the environment advances by the chosen horizon, and the agent is rewarded when the price confirmed its call with a large enough move and penalised when the price went the other way. Staying neutral earns a small reward when the price did nothing and a penalty when it missed a move it should have called. Over many passes through the history, the learning algorithm adjusts the network towards the calls that were rewarded.

The algorithm matters. FundSpec's builder offers PPO, which clips each policy update to keep learning stable, A2C, which runs deterministic updates across several environments at once, and DQN, which applies Q learning to the discrete set of actions. The reward thresholds matter more: the trader profile sets how large a move counts as confirming a call, so a conservative model demands a decisive move before it signals and an aggressive one acts on tighter thresholds.

Testing on history the model never saw

A model that has memorised its training data will look brilliant on that data and fail on anything new. Every FundSpec model is scored on a slice of history it never trained on, and only the out of sample result counts. A model that does not clear the bar on unseen data is not released.

Even out of sample testing has limits. History contains a finite number of regimes, and a model trained through a bull market has not been tested through a crash. The released record, described next, is the only test that continues.

The published record

Once a model is released, every call it makes during the live session is recorded, and every closed trade is added to its record. The record shows the number of closed trades on the ticker, how often the model won, the average size of a winning move, and the profit factor, total gains divided by total losses. It shows the most recent winning trade with its direction, entry date, gain and holding period. Losses sit in the same list as wins.

The record is computed from the model's own published calls against the prices that followed, so it cannot be curated after the fact. It is also history. A model with a strong record has behaved well in the conditions it has faced; it has made no promise about the next ones.

Reading a record honestly

Three questions separate a useful record from a flattering one. How many trades is it built on? A win rate from ten trades tells you almost nothing; from a hundred it starts to mean something. What is the profit factor, not just the win rate? A model that wins often with small gains and loses rarely with large ones can have a high win rate and lose money. How long is the holding period? A model whose average trade lasts three sessions is making a different kind of call from one that holds for weeks, and its record should be judged against the costs of trading that often.

Training your own

Subscribers can train a model on any US stock or ETF. The builder walks through the choices: the ticker, daily or intraday bars, the algorithm, the trader profile, how many bars the network sees per decision, the reward horizon, how much history to train on and how long to train. The guided default explores architectures automatically; the premium tiers allow deeper networks and manual tuning. Training takes hours, the model is tested out of sample like every other, and it then publishes signals to you the same way.

What the models do not do

They do not place trades. FundSpec has no brokerage connection and executes nothing; the models publish calls and the record of what followed, and every decision to act is yours. They do not read news, filings or fundamentals. And they do not guarantee anything: a track record is evidence about the past, and the honest way to use it is as one input among the others FundSpec puts on the same page.

How FundSpec shows it

The Day Trading Models screen lists the released models with their live calls for the session and their records. Everyone can see the recent winning trades the models have closed; FundSpec Unlimited shows the calls as they are published. Each model's overview page carries the full performance record. Subscribers train their own models from the My Models screen.