A quantitative research platform for implementing, backtesting, comparing, optimizing, and validating systematic trading strategies on historical equity data.
Built with Python, pandas, NumPy, Plotly, Streamlit and yfinance.
This project is designed for quantitative strategy research and education. It is not a trading signal service, automated execution system, or source of investment recommendations.
The platform is built around a simple idea: a strong historical backtest is only the beginning of strategy research.
A strategy may perform well because its underlying idea captured persistent market behavior, because its parameters happened to fit a particular historical period, or simply because the selected sample was favorable. The platform therefore goes beyond individual backtests by providing tools for parameter sensitivity analysis, robustness diagnostics, out-of-sample testing, and walk-forward validation.
The application is organized around a progressively more demanding research process:
Strategy implementation
↓
Historical backtest
↓
Cross-strategy comparison
↓
Parameter optimization
↓
Robustness diagnostics
↓
Out-of-sample validation
↓
Walk-forward validation
Each stage answers a different question.
Backtesting
How did this strategy behave historically under the selected assumptions?
Comparison
How did different systematic approaches behave over the same market period?
Optimization
Which tested parameters maximize the selected historical objective?
Robustness analysis
Is the mathematical optimum surrounded by similarly strong parameter choices, or is it an isolated peak?
Out-of-sample validation
Do parameters selected on earlier observations retain their behavior on data that was not used to select them?
Walk-forward validation
Does parameter selection continue to work when the optimization and unseen-test process is repeated through time?
The platform deliberately separates these questions rather than treating the highest in-sample performance as sufficient evidence of strategy quality.
The screenshots below use AAPL over a common historical period to demonstrate the main research workflow.
They are intended to show how the application moves from strategy comparison and individual backtesting toward parameter research, robustness analysis, and repeated out-of-sample validation.
Compare all registered strategies under the same ticker, time period, capital, transaction-cost, and risk-free-rate assumptions.
Inspect one strategy's parameters, headline metrics, market exposure, trades, and strategy-specific indicators.
The example below shows the default Donchian Breakout implementation using prior rolling High/Low channels.
Compare the default configuration with the mathematical optimum and inspect how a broad parameter search is refined locally around the most promising region.
The optimization surface is used to investigate whether strong historical performance is concentrated around a narrow parameter peak or supported by a wider near-optimal region.
Repeated chronological training and unseen-test folds evaluate whether parameter selection remains useful outside the observations used for optimization.
Walk-forward validation also records how selected parameters change across folds and compounds the unseen test periods into continuous stitched out-of-sample research curves.
Run every registered strategy on the same ticker and historical period under common assumptions.
The comparison framework reports metrics including:
- Final portfolio value
- Total return
- CAGR
- Annualized volatility
- Sharpe ratio
- Maximum drawdown
- Rebalancing activity
- Cumulative one-way turnover
- Average market exposure
- Buy and Hold benchmark performance
Strategies can be ranked across multiple performance and risk dimensions rather than by return alone.
The comparison page can also optionally use optimized parameters. Because those parameters are selected from the same historical sample being compared, optimized comparison results are explicitly treated as in-sample research rather than validated evidence.
Study one strategy in greater detail.
The Single Strategy page provides:
- Configurable strategy parameters
- Strategy-specific indicators
- Price and signal visualizations
- Position behavior
- Strategy and Buy and Hold equity curves
- Drawdown analysis
- Performance metrics
- Rebalancing activity
- One-way turnover
- Average market exposure
An optional optimized-parameter mode can search the strategy's predefined research grid before running the backtest.
Optimized results are clearly identified as in-sample and potentially vulnerable to overfitting.
The Optimization Lab studies how strategy performance changes across parameter configurations.
It supports:
- Predefined research grids
- Custom parameter grids
- Coarse-to-fine parameter search
- Multiple optimization objectives
- Ranked parameter combinations
- Two-dimensional parameter sensitivity heatmaps
- Higher-dimensional parameter sensitivity profiles
- Boundary-optimum diagnostics
- Near-optimal-region analysis
- Stability-oriented parameter selection
Supported optimization objectives include:
- Sharpe ratio
- Total return
- CAGR
- Annualized volatility
- Maximum drawdown
The platform distinguishes between the raw optimum and a stability-oriented candidate.
The raw optimum is the mathematical best tested parameter combination for the selected objective.
The stability-oriented candidate is selected from the near-optimal region using the behavior of its surrounding tested parameter neighborhood. This provides an alternative to automatically trusting a single isolated optimum.
A high-performing parameter combination is not necessarily a robust one.
The platform therefore analyzes the shape of the tested optimization surface.
Diagnostics include:
- Share of tested combinations within 5% of the optimum
- Share within 10% of the optimum
- Performance of the immediate tested-grid neighborhood
- Median nearby objective degradation
- Boundary-optimum detection
- Stability-oriented candidate selection
Optimization surfaces are summarized using descriptive labels:
- Broad plateau
- Moderate plateau
- Narrow peak
A broad near-optimal region can be more reassuring than an isolated optimum because small parameter changes do not immediately destroy the historical result.
These classifications are intentionally treated as descriptive heuristics, not statistical proof that a strategy is or is not overfit.
Optimization tells us what worked best on historical training data.
Validation asks the more important question:
What happened after those parameters were selected?
The Validation Lab provides both single-split and walk-forward out-of-sample testing.
The historical sample is divided into two chronological regions:
Training period
↓
Parameter optimization
↓
Parameters frozen
↓
Unseen test period
Observations before the split are used for optimization.
Observations on or after the split are reserved for out-of-sample evaluation.
The test period is not used to select the raw optimum or stability-oriented candidate.
The platform compares:
- Default parameters
- Raw training optimum
- Stability-oriented candidate
- Buy and Hold
Training and test performance are displayed separately so deterioration outside the optimization sample can be observed directly.
Expanding walk-forward validation repeats the optimization and unseen-test process through time while retaining all previous training history.
For example:
2020–2021 → test 2022
2020–2022 → test 2023
2020–2023 → test 2024
2020–2024 → test 2025
Each test window remains unseen when its parameters are selected.
This evaluates whether parameter selection remains useful as more historical information becomes available.
Rolling validation instead keeps a fixed-length recent training window.
For example:
2020–2021 → test 2022
2021–2022 → test 2023
2022–2023 → test 2024
2023–2024 → test 2025
Older observations eventually leave the optimization sample.
This allows the platform to investigate whether strategy behavior appears more stable when parameter selection emphasizes relatively recent market regimes.
For every fold, the platform records:
- Training period
- Test period
- Training robustness
- Default test objective
- Raw-optimum test objective
- Robust-candidate test objective
- Buy and Hold test objective
- Out-of-sample degradation
- Whether optimized parameters beat the defaults
- Whether the stability-oriented candidate beat the raw optimum
- Selected parameter history
Repeated unseen periods are also summarized into aggregate walk-forward evidence.
The application builds stitched out-of-sample equity curves to show how the sequence of unseen test periods behaves as a continuous research portfolio.
Fold-level metrics remain independently rebased for comparison, while stitched strategy curves transition continuously between successive out-of-sample folds. Transaction costs at fold boundaries therefore apply to the actual change in exposure rather than unnecessarily liquidating and rebuilding an unchanged position.
Repeated out-of-sample success is stronger evidence than a single optimized historical backtest, but it still does not guarantee future performance.
The Strategy Guide documents the ideas represented by the platform rather than treating strategies as black boxes.
Each registered strategy includes information such as:
- Strategy family
- Strategy subtype
- Signal style
- Main features
- Core idea
- Strategy thesis
- Market behavior it is designed around
- Favorable environments
- Unfavorable environments
- Main risks
- Parameter definitions
- Optimization considerations
- Validation considerations
- Related strategies
Descriptions of favorable and unfavorable environments refer to market behavior, not claims that a strategy will reliably outperform on a particular stock, sector, or future period.
The Strategy Guide is intended to connect the mathematical implementation with the economic intuition behind each research rule.
The platform currently includes 15 strategies covering several systematic research ideas.
- Moving-Average Crossover
- EMA Crossover
- MACD Crossover
- ADX Trend Filter
- Volatility-Filtered Trend
- Daily and Weekly Trend Confirmation
- Donchian Breakout
- ATR Breakout
- Keltner Channel Breakout
- Price Momentum
- Volatility-Targeted Momentum
- RSI Mean Reversion
- Bollinger-Band Mean Reversion
- Stochastic Mean Reversion
- CCI Mean Reversion
The Strategy Guide provides the detailed interpretation, parameter definitions, risks, and intended market behavior for each implementation.
The number of strategies is intentionally limited.
The goal of the platform is not to maximize the size of the strategy catalog. Instead, the included strategies are intended to represent meaningfully different systematic ideas while keeping the research environment understandable, auditable, and extensible.
Many apparently different strategies are better understood as parameter variations of the same underlying idea.
For example, separate moving-average window combinations do not need to become separate strategies. They belong inside the parameter and optimization framework.
Adding dozens of minor indicator variations would increase the feature count without necessarily increasing the research value of the project.
The included set therefore prioritizes:
- Coverage of distinct systematic ideas
- Interpretability
- Parameter research
- Robustness analysis
- Out-of-sample validation
- Code clarity
- Extensibility
The 15 included strategies should be viewed as a curated research set and examples of the framework, not as the limits of the platform.
The strategy architecture is designed to be extensible.
A new strategy can be integrated into the existing research workflow without creating separate comparison, optimization, or validation systems.
Add a new Python file inside:
strategies/
For example:
strategies/my_strategy.py
A long-or-cash strategy should generate a signal column:
import pandas as pd
def generate_my_strategy_signals(
data: pd.DataFrame,
window: int = 20,
) -> pd.DataFrame:
result = data.copy()
# Calculate indicators here.
result["signal"] = ...
return resultThe signal convention is:
0 = cash
1 = long
Strategies using continuous position sizing can additionally provide:
result["target_position"]The backtester prioritizes target_position when it is available, allowing fractional exposure rather than forcing every strategy into binary long/cash positions.
Import and register the implementation in:
strategies/__init__.py
The central strategy registry connects the implementation to the rest of the application.
Add the strategy's metadata and parameter definitions to the central metadata registry.
Metadata powers the Strategy Guide and other research interfaces, including information such as:
- Display name
- Family
- Subtype
- Features
- Signal style
- Description
- Thesis
- Favorable environment
- Unfavorable environment
- Risks
- Parameter specifications
- Optimization guidance
- Validation guidance
Automated tests verify that registered strategies contain the required documentation fields.
Define a predefined research grid for the strategy so it can participate in Optimization Lab and optimized research workflows.
Parameter grids should be broad enough to investigate meaningful behavior without creating unnecessarily large brute-force searches.
After adding a strategy:
python -m pytestThe existing test suite checks strategy output, registry integrity, metadata completeness, backtester compatibility, and other research assumptions.
Historical market data is downloaded using yfinance.
Adjusted historical prices are used so corporate actions represented by the adjusted series are reflected in the backtest.
Daily simple returns are calculated from adjusted closing prices.
Strategy indicators and signals may use information available on trading day (t).
The resulting desired exposure is shifted by one trading observation before portfolio returns are calculated.
Conceptually:
Information available on day t
↓
Signal calculated on day t
↓
Position used on day t + 1
This prevents a strategy from using a closing-price signal to retroactively earn the return that produced that same closing price.
Breakout strategies may additionally use prior-period channel levels where required by their definitions.
The platform is currently long-only.
Most strategies use binary exposure:
0 = cash
1 = fully invested
Strategies designed around risk-managed position sizing can use continuous target exposure.
Negative exposure is rejected by the backtester.
Transaction costs are modeled as configurable one-way proportional costs on changes in portfolio exposure.
For example:
0.0 → 1.0 = 1.0 one-way turnover
1.0 → 0.0 = 1.0 one-way turnover
0.4 → 0.7 = 0.3 one-way turnover
This allows the same transaction-cost model to support both binary and continuously sized strategies.
Buy and Hold is treated as an executable benchmark rather than a cost-free theoretical curve.
The benchmark pays one initial one-way transaction cost when its position is established and then remains invested.
Average market exposure is calculated from the realized position series.
For binary strategies, it approximately represents the proportion of the backtest spent invested.
For continuously sized strategies, it represents the average fraction of capital exposed to the asset.
This provides useful context when comparing a partially invested strategy with a 100%-exposed Buy and Hold benchmark.
The platform reports several complementary performance measures.
Total compounded return over the complete research period.
Compound annual growth rate based on elapsed calendar time.
Daily return volatility annualized using 252 trading days.
Annualized risk-adjusted return relative to the selected annual risk-free rate.
Largest historical peak-to-trough decline in portfolio equity.
Number of trading observations on which portfolio exposure changes.
Cumulative absolute change in portfolio exposure.
Mean realized portfolio position over the research period.
No single metric is treated as sufficient evidence of strategy quality.
Parameter optimization evaluates combinations from a defined parameter grid under identical backtesting assumptions.
The selected objective determines how parameter combinations are ranked.
Higher values are preferred for:
- Sharpe ratio
- Total return
- CAGR
- Maximum drawdown, where values closer to zero are better
Lower values are preferred for:
- Annualized volatility
Optimization results are historical and in-sample unless they are subsequently evaluated through the Validation Lab.
For strategies with larger parameter spaces, the Optimization Lab supports a two-stage search.
A relatively sparse grid explores the wider parameter space.
A denser grid is constructed around the broad-stage winner.
The refinement process interpolates candidate values inside the local neighborhood rather than simply adding a fixed numeric amount to every parameter.
Integer parameters remain integer-valued, metadata bounds are respected, and invalid parameter relationships are filtered before evaluation.
This provides greater local resolution without requiring exhaustive evaluation of every possible numeric value.
The project deliberately distinguishes optimization from robustness.
Optimization asks:
Which tested parameter combination produced the best historical objective?
Robustness asks:
How dependent was that result on selecting exactly those parameters?
An isolated optimum surrounded by substantially worse configurations may indicate greater parameter-selection risk than a broad region of similarly strong results.
The stability-oriented candidate therefore provides an alternative research reference point, but it is not automatically assumed to be superior to the raw optimum.
Its usefulness must still be evaluated out of sample.
The project intentionally avoids treating optimization as the final research result.
A parameter configuration can perform extremely well on the observations used to select it and poorly afterward.
For this reason, the research workflow separates:
Parameter selection
from:
Parameter evaluation
Single-split validation provides one unseen historical experiment.
Walk-forward validation repeats that experiment through time.
Neither eliminates overfitting or guarantees future performance, but both provide more meaningful evidence than reporting optimized in-sample results alone.
A simplified view of the repository:
Quant-Strategy-Research-Platform/
│
├── app.py
├── main.py
│
├── backtester.py
├── metrics.py
├── comparison.py
├── optimization.py
├── validation.py
├── plots.py
│
├── comparison_app.py
├── single_strategy_app.py
├── optimization_app.py
├── validation_app.py
├── strategy_guide_app.py
│
├── run_optimization.py
│
├── strategies/
│ ├── __init__.py
│ ├── moving_average.py
│ ├── ema_crossover.py
│ ├── rsi.py
│ ├── bollinger.py
│ ├── stochastic.py
│ ├── macd.py
│ ├── momentum.py
│ ├── donchian.py
│ ├── atr_breakout.py
│ ├── adx.py
│ ├── cci.py
│ ├── keltner.py
│ ├── volatility_filter.py
│ ├── multi_timeframe_trend.py
│ └── volatility_targeted_momentum.py
│
├── tests/
│ ├── test_backtester.py
│ ├── test_comparison.py
│ ├── test_metrics.py
│ ├── test_optimization.py
│ ├── test_optimization_robustness.py
│ ├── test_strategies.py
│ ├── test_strategy_guide_metadata.py
│ └── test_validation.py
│
├── assets/
│ ├── comparison-overview.png
│ ├── single-strategy-donchian.png
│ ├── optimization-summary.png
│ ├── optimization-robustness.png
│ ├── validation-walk-forward-overview.png
│ └── validation-walk-forward-stability.png
│
├── LICENSE
├── requirements.txt
├── .gitignore
└── README.md
Handles:
- Historical data preparation
- Daily returns
- Signal-to-position timing
- Long-only exposure
- Transaction costs
- Strategy returns
- Buy and Hold benchmark
- Equity curves
- Drawdowns
Calculates:
- Total return
- CAGR
- Annualized volatility
- Sharpe ratio
- Maximum drawdown
Runs registered strategies under common assumptions and builds cross-strategy rankings.
Provides:
- Parameter-grid optimization
- Multiple objectives
- Optimization gain calculations
- Coarse-to-fine refinement
- Parameter sensitivity analysis
- Robustness diagnostics
- Stability-oriented candidate selection
Provides:
- Strict chronological train/test validation
- Frozen-parameter evaluation
- Out-of-sample degradation analysis
- Expanding walk-forward validation
- Rolling walk-forward validation
- Parameter history
- Fold-level evidence
- Continuous stitched out-of-sample equity analysis
Acts as the central strategy registry.
It connects:
- Strategy implementations
- Default parameters
- Strategy metadata
- Parameter specifications
- Predefined optimization grids
This registry-driven structure allows the research pages to work across strategies without requiring separate application logic for every implementation.
The project is built primarily with:
- Python
- Streamlit
- pandas
- NumPy
- Plotly
- yfinance
- pytest
Clone the repository:
git clone https://github.com/enesozs/Quant-Strategy-Research-Platform.git
cd Quant-Strategy-Research-PlatformCreate a virtual environment:
python -m venv .venvActivate it on macOS/Linux:
source .venv/bin/activateOn Windows:
.venv\Scripts\activateInstall dependencies:
pip install -r requirements.txtRun the application:
python -m streamlit run app.pyRun the complete automated test suite with:
python -m pytestThe test suite covers areas including:
- Signal validity
- Strategy registry behavior
- Strategy metadata completeness
- Look-ahead protection
- Position timing
- Transaction costs
- Fractional exposure
- Equity compounding
- Performance metrics
- Optimization direction
- Invalid parameter combinations
- Coarse-to-fine refinement
- Optimization robustness
- Strict train/test separation
- Frozen optimized parameters
- Expanding walk-forward validation
- Rolling walk-forward validation
- Stitched out-of-sample behavior
At the time of the initial repository release, the project passes 192 automated tests.
This platform intentionally uses a simplified research model.
Important limitations include:
Historical performance does not guarantee future performance.
A successful backtest demonstrates what would have happened under the modeled assumptions, not what will happen in the future.
Historical market data is obtained through yfinance.
Data availability, adjustments, corrections, and upstream provider behavior are outside the control of this project.
Uninvested capital is assumed to earn zero return.
The selected risk-free rate is used for Sharpe-ratio calculations but is not credited to cash balances.
Transaction costs are modeled as proportional one-way costs on changes in exposure.
The model does not explicitly simulate:
- Bid/ask spreads
- Market impact
- Order-book liquidity
- Variable commissions
- Execution latency
Where a strategy permits exposure above 100%, the additional exposure represents simplified leverage.
Financing costs, borrowing constraints, margin requirements, and liquidation mechanics are not modeled.
The current backtester is long-only and does not support negative portfolio exposure.
Historical optimization can overfit.
The raw optimum should not be interpreted as the universally best parameter configuration.
Broad-plateau, moderate-plateau, narrow-peak, neighborhood, and stability-oriented diagnostics are descriptive research heuristics.
They are not statistical guarantees against overfitting.
Out-of-sample testing reduces direct parameter-selection bias because test observations are not used to choose the evaluated parameters.
However, researchers can still introduce bias through repeated experimentation, strategy selection, universe selection, date selection, and interpretation.
Walk-forward analysis provides repeated unseen historical tests and is more demanding than a single optimized backtest.
It still cannot establish future profitability.
The current project focuses on researching one underlying asset at a time.
It is not a multi-asset portfolio-construction or portfolio-optimization engine.
The application does not connect to a broker and does not execute trades.
This project is intended for:
- Quantitative strategy research
- Historical experimentation
- Studying systematic trading concepts
- Parameter sensitivity analysis
- Learning about overfitting
- Comparing systematic approaches
- Practicing out-of-sample validation
- Exploring walk-forward methodology
- Extending the framework with new strategy ideas
It is not intended to provide personalized investment advice or live trading recommendations.
The project intentionally prioritizes:
Transparency over black-box complexity
Strategy rules and parameter choices should be inspectable.
Research workflow over headline backtest performance
Optimization is treated as an intermediate step rather than the final answer.
Validation over parameter chasing
Unseen historical data is used to challenge conclusions formed during optimization.
Interpretability over strategy count
A curated set of understandable strategies is preferred to a large catalog of minor variations.
Extensibility over exhaustiveness
The included strategies demonstrate the framework; users can add their own research rules without redesigning the entire application.
Honest limitations over unrealistic precision
The platform explicitly documents assumptions that are not modeled rather than implying that a historical simulation is equivalent to real-world execution.
The current feature set is intentionally focused on strategy-level research.
Possible future extensions could include:
- Additional user-defined strategy families
- More advanced transaction-cost models
- Alternative market-data providers
- Additional out-of-sample research techniques
- Multi-asset research
- Portfolio-level analysis
- Short-selling support
- Financing-cost modeling
These are deliberately treated as extensions rather than requirements for the current research platform.
This software is provided for educational and research purposes only.
Nothing in this repository or application constitutes financial, investment, trading, legal, or tax advice.
Historical simulations, optimized parameters, robustness diagnostics, and out-of-sample results do not guarantee future performance.
Any use of the software or its outputs for real-world financial decisions is entirely at the user's own risk.