What is StockVisionz?

StockVisionz is a user-driven algorithmic trading research platform for US equities. You search any ticker, the platform fetches and stores the data automatically, and you trigger backtests or ML model analysis on demand — all from the dashboard.

The platform is built for two audiences: retail traders who want to test rule-based strategies with professional-grade metrics, and quant researchers who want to compare ML model families under rigorous, apples-to-apples conditions.

Running Your First Backtest

The core workflow is deliberately simple. Everything happens from a single dashboard.

• —Open the dashboard and type any US equity ticker symbol into the search bar — for example, AAPL, MSFT, or NVDA.
• —If the ticker is not yet in the database, the platform fetches price history automatically, computes all technical indicators, and stores everything. This takes a few seconds.
• —Select a strategy from the dropdown. The default strategy is SMA 50/200 crossover (Golden Cross / Death Cross).
• —Select a date range using the range selector.
• —Click Run Backtest. Results populate in the dashboard within seconds.
• —Scroll down to view the full tearsheet — equity curve, drawdown chart, monthly returns heatmap, and all key metrics.

Reading Your Results

After a backtest completes, four metric cards appear at the top of the dashboard. Here is what each one means:

The equity curve shows portfolio value over time normalized to a starting value of 100. The drawdown chart directly below it shows how far the portfolio was underwater at each point in time. A flat drawdown chart means the strategy recovered losses quickly.

Survivorship Bias

Survivorship bias is one of the most common sources of inflated backtest results in retail trading research, and it is important to understand how it affects StockVisionz in beta.

The current data source — yfinance — only returns price data for companies that are currently trading. If you backtest a strategy over the past ten years, you are implicitly backtesting only on companies that survived that entire period. Every company that went bankrupt, was delisted, or was acquired disappears from the data. This makes strategies appear more profitable than they would have been in real-time deployment.

The platform architecture is designed to make this easy to fix. All data fetching is abstracted behind a DataAdapter interface. Swapping from yfinance to a survivorship-bias-free vendor like Polygon.io — which includes delisted company history — is a single configuration change. This upgrade is planned for the production release.

Data Leakage

Data leakage is when information from the future is accidentally used to generate a signal in the past. It is the most common reason a backtest looks exceptional but performs poorly in live trading.

The most common form in daily equity backtesting: a strategy generates a buy signal using today's closing price and then enters the trade at today's closing price. In reality, by the time you know the closing price, the market is already closed. The trade can only execute at the next day's open.

In the ML layer, data leakage takes a subtler form: features that include same-day closing data when predicting that same day's return. Every feature in StockVisionz is lagged at least one period relative to the prediction target. This is enforced in the feature matrix builder and verified explicitly before any model result is reported.

Walk-Forward Validation

A simple train/test split — train on 2015-2020, test on 2021-2022 — is the standard academic approach, but it has a critical flaw for financial ML: it only tests one time period. A model that happened to be trained before a bull market and tested during one will look great. The same model tested during a crash will look terrible.

Walk-forward validation approximates how a model actually gets deployed. The training window rolls forward through time, and the model is evaluated on the immediately following out-of-sample period — over and over across the full history. The result is a distribution of out-of-sample performance across many market regimes, not a single lucky or unlucky test window.

How StockVisionz implements it:

• —A training window of fixed length is placed at the start of the available history.
• —The model trains on that window only.
• —It is evaluated on the next N periods it has never seen.
• —The window advances forward. The model retrains. The process repeats.
• —Out-of-sample signals across all windows are stitched together and run through the backtester to compute Directional Profitability.

Meta-Labeling

Predicting next-day market direction directly is a notoriously hard problem. Most models flatline around 51-53% out-of-sample accuracy — barely better than a coin flip. This is not a failure of the model; it reflects the genuine difficulty of the problem.

Meta-labeling reframes the ML model's role entirely. Instead of asking 'will the market go up or down tomorrow?', the model asks a narrower, more tractable question: 'given that my rule-based strategy just generated a buy signal, is this particular signal likely to be profitable?'

The four-step flow:

• —A rule-based strategy (e.g. SMA crossover) generates a primary buy or sell signal.
• —The ML model receives that signal and the current feature state as inputs.
• —The ML model outputs a confidence score between 0 and 1.
• —High confidence: the trade executes. Low confidence: the trade is filtered out regardless of the primary signal.

Meta-labeled models typically show lower trade frequency — more signals are filtered — but higher precision on the trades they pass through. The key success metric is Sharpe ratio improvement over the unfiltered rule-based strategy, not raw classification accuracy.

Concept Drift

A model trained during a low-volatility bull market will learn patterns specific to that regime. When the market shifts — a rate hike cycle, a crash, a sector rotation — the patterns change and model performance degrades silently. This is concept drift.

StockVisionz monitors concept drift by tracking feature importance rankings over time. When the features that were most predictive during training are no longer the most predictive in recent data, that is a signal that the model needs retraining. Feature importance history is stored in the database per model run and surfaced in the dashboard comparison view.

Available Strategies

Each strategy is a self-contained module. It accepts the indicator DataFrame for the selected ticker and returns a signal series. New strategies can be added without modifying the backtesting runner.

Technical Indicators

All indicators are pre-computed from price data and stored in the database. They are never recalculated at query time. Every consumer — the backtester, the ML feature builder, and the dashboard — reads from the same stored values, guaranteeing consistency.

Understanding the Tearsheet

Every backtest automatically generates a full tearsheet. Here is what each section shows and how to interpret it.

Equity Curve

Portfolio value over time, normalized to 100 at the start. A steadily rising curve with shallow drawdowns is ideal. Sudden steep drops indicate periods where the strategy was significantly underwater.

Underwater Chart

Shows the percentage decline from the most recent equity peak at each point in time. The deeper the trough, the worse the drawdown. The longer the chart stays below zero, the longer the recovery period.

Monthly Returns Heatmap

Each cell shows the return for a specific month and year. Green is positive, red is negative. The heatmap makes seasonality and bad-year clustering immediately visible.

Risk Metrics

How Models Are Compared

All model families are evaluated under identical conditions per comparison group: same ticker, same feature set, same training and test windows. No model receives a home field advantage.

Shared feature set:

• —Lagged returns — 1, 2, 5, and 10 day
• —RSI (14-period)
• —MACD signal line
• —ATR (14-period)
• —Volume ratio (current volume / 20-day volume SMA)
• —Bollinger Band position (where price sits relative to the bands)

Prediction target:

Binary classification — 1 if next-day return is positive, 0 if negative. All supervised models share this target. RL agents learn a trading policy instead and are not evaluated on classification metrics.

Logistic Regression

Logistic Regression is the simplest supervised model in the platform. It fits a linear decision boundary in feature space and outputs a probability score for the positive class.

Its value is interpretability and speed, not raw predictive power. It establishes a baseline that more complex models must beat to justify their added complexity. If XGBoost barely outperforms Logistic Regression on your ticker, that is useful information — the additional complexity of a tree ensemble is not warranted.

XGBoost

XGBoost is the primary ML baseline on StockVisionz. It is a gradient-boosted tree ensemble — it builds a sequence of decision trees where each tree corrects the errors of the previous one.

XGBoost consistently outperforms other tree methods on tabular financial data and trains significantly faster than Random Forest at scale. All XGBoost instances on StockVisionz are configured with GPU acceleration (tree_method='hist', device='cuda').

Overfitting risk:

XGBoost can overfit aggressively if hyperparameters are not regularized. Walk-forward validation and hyperparameter search confined to the training window are both required before trusting any XGBoost result. Feature importance scores are logged per run to monitor drift.

LSTM

The LSTM (Long Short-Term Memory) is a recurrent neural network architecture designed to learn patterns in sequential data. Unlike tree models that receive a flat feature vector, the LSTM receives a 20-day lookback window — a sequence of daily feature states — and learns temporal patterns across that sequence.

Architecture:

• —Input shape: (batch_size, 20 days, number of features)
• —1-2 LSTM hidden layers with 64-128 units
• —Dropout of 0.2-0.4 for regularization
• —Fully connected output layer with sigmoid activation
• —Binary Cross-Entropy loss, Adam optimizer with learning rate scheduler
• —Early stopping on validation loss

Reinforcement Learning Agents

The RL layer is the most experimental component of StockVisionz. Three agents — DQN, PPO, and A2C — learn a trading policy by interacting with a simulated trading environment built from the stored price and indicator data.

Environment design:

• —State: current indicator values — same feature set as supervised models
• —Action space: discrete — 0 = Hold, 1 = Buy, 2 = Sell
• —Reward: risk-adjusted return per step, Sharpe-based, penalizes drawdown and excessive turnover
• —Episode: one contiguous trading period from the training window

General

Why is my ticker missing from the search results?

StockVisionz fetches data on demand. If a ticker is not yet in the database, simply search for it — the platform will fetch price history automatically, compute indicators, and have it ready for analysis within a few seconds. The ticker must be a valid US equity symbol currently trading.

How often is price data updated?

Price data for active tickers refreshes automatically each trading day after market close via a scheduled job. If you are viewing a ticker that has not been searched recently, the data may be a day behind. Triggering any analysis on that ticker will pull the latest available data.

What does the survivorship bias warning mean?

The current data source only includes companies currently trading. Companies that were delisted, went bankrupt, or were acquired are not included in historical data. This means backtests will appear more profitable than they would have been in real time. See the Survivorship Bias section in Platform Concepts for the full explanation.

Is this live trading or paper trading?

StockVisionz is currently a research and backtesting platform. Paper trading via Alpaca is planned for a future phase. No real money is involved at any stage of the current platform.

Why does my ML model take longer than my backtest?

Backtests are vectorized and run in seconds. ML training — especially walk-forward validation, which trains the model multiple times across the full history — takes longer. XGBoost and LSTM training run on GPU, which dramatically reduces training time compared to CPU. If you are seeing very long training times, confirm that GPU acceleration is active in the platform settings.

Why are my ML accuracy scores only around 51-53%?

This is expected and normal. Predicting next-day market direction is a genuinely hard problem. The market is not random, but it is close enough to random that 51-53% accuracy on a binary classification task is a realistic out-of-sample result. The more meaningful metric is Directional Profitability — whether running the model's signals through the backtester produces a positive Sharpe ratio. A model with 53% accuracy and a Sharpe of 1.2 is a good model.