Financial AI Tasks
Comprehensive overview of the four core financial AI tasks supported by FinWorld
Task Overview
FinWorld supports four core financial AI tasks that cover the entire spectrum of financial AI applications. Each task is designed with specific mathematical formulations and can be approached using multiple AI paradigms including traditional ML, deep learning, reinforcement learning, and large language models.
Notation
Throughout this documentation, we use the following notation:
- Single Asset: $\mathbf{x}_{1:T} = \{x_1, ..., x_T\} \in \mathbb{R}^{T \times D}$ - Historical endogenous time series
- Multi-Asset: $\mathbf{x}_{1:T} \in \mathbb{R}^{N \times T \times D}$ - Historical price sequences for $N$ assets
- Exogenous Variables: $\mathbf{z}_{1:T_{\mathrm{ex}}} = \{\mathbf{z}_1, ..., \mathbf{z}_{T_{\mathrm{ex}}}\} \in \mathbb{R}^{T_{\mathrm{ex}} \times C}$ - Technical indicators, financial factors, news sentiment
Time Series Forecasting
Financial time series forecasting differs from traditional forecasting in two key ways: it predicts returns rather than prices (due to non-stationarity), and it typically involves multiple assets simultaneously to reflect real-world portfolio management requirements.
Mathematical Formulation
Given: Historical prices of $N$ stocks $\mathbf{x}_{1:T} \in \mathbb{R}^{N \times T \times D}$ and exogenous variables $\mathbf{z}_{1:T_{\mathrm{ex}}}$
Goal: Predict future $S$-day relative returns for all assets
$$\hat{\mathbf{R}}_{T+1:T+S} = \mathcal{F}_\theta(\mathbf{x}_{1:T}, \mathbf{z}_{1:T_{\mathrm{ex}}})$$
Where:
- $\hat{\mathbf{R}}_{T+1:T+S} \in \mathbb{R}^{N \times S}$ are predicted relative returns
- $\mathbf{R}_t = \frac{\mathbf{x}_t}{\mathbf{x}_T} - 1$ (relative return calculation)
- $\mathcal{F}_\theta$ is the forecasting model parameterized by $\theta$
🔮 Multi-Asset Time Series Forecasting Workflow
1
Historical Price Data
Input: $\mathbf{x}_{1:T} \in \mathbb{R}^{N \times T \times D}$
📈 AAPL
30-day history
💻 MSFT
30-day history
🔍 GOOGL
30-day history
⚡ TSLA
30-day history
🏦 JPM
30-day history
2
Feature Engineering
Processing + Exogenous Variables
🔧 Price Features
OHLCV + Returns
📊 Technical Indicators
15 indicators
🌍 Exogenous Variables
$\mathbf{z}_{1:T_{\mathrm{ex}}}$
📈 Time Window
T=30 days
3
Multi-Asset Forecasting Model
Model: $\hat{\mathbf{R}}_{T+1:T+S}$
S=5
Forecast Horizon (days)
Model
$\mathcal{F}_\theta$
Architecture
LSTM
Output
$\mathbb{R}^{N \times S}$
4
Predicted Relative Returns
Output: $\mathbf{R}_{t} = \frac{\mathbf{x}_t}{\mathbf{x}_T} - 1$
📈 AAPL
+2.3%
💻 MSFT
+1.8%
🔍 GOOGL
+1.5%
⚡ TSLA
+3.2%
🏦 JPM
+1.9%
5
Portfolio Performance
Task
Expected Return
+2.2%
Sharpe Ratio
1.8
Volatility
12-15%
Confidence
78%
Key Features
- Multi-Asset Prediction - Simultaneous forecasting for multiple assets
- Return-Based - Predicts returns rather than absolute prices
- Exogenous Integration - Incorporates technical indicators and external factors
- Flexible Horizons - Supports various prediction horizons (S days)
Algorithmic Trading
Algorithmic trading involves the design, simulation, and evaluation of systematic trading strategies with emphasis on real-time decision making and risk management. FinWorld supports both ML/DL-based and RL-based approaches.
ML/DL-based Approach
Goal: Predict future price returns or movement directions
$$\hat{y}_{T+1:T+S} = \mathcal{F}_\theta(\mathbf{x}_{1:T}, \mathbf{z}_{1:T_{\mathrm{ex}}})$$
Trading Actions:
- Buy if $\hat{y}_t > \tau$
- Sell if $\hat{y}_t < -\tau$
- Hold otherwise
Where:
- $\hat{y}_t$ represents predicted return, probability, or trading signals
- $\tau$ is a threshold parameter
- Actions are determined by pre-defined decision rules
RL-based Approach
Modeled as Markov Decision Process (MDP):
At each time step $t$:
- Observe state $s_t$ (constructed from $\mathbf{x}_{1:t}$ and $\mathbf{z}_{1:t}$)
- Choose action $a_t \in \mathcal{A}$
- Earn reward $r_t$
- Transition to next state $s_{t+1}$
Objective: Learn policy $\pi_\theta(a_t \mid s_t)$ that maximizes:
$$\max_\theta~ \mathbb{E}_{\pi_\theta} \left[\sum_{t=T+1}^{T+S} r_t\right]$$
🤖 Algorithmic Trading Workflow
📊 ML/DL-based Approach
1
Market Data
Input: $\mathbf{x}_{1:T} \in \mathbb{R}^{T \times D}$
📈 AAPL
$150.25
📊 Volume
2.1M
📰 News
iPhone sales up
⏰ Time
09:30:15 EST
2
Feature Engineering
Processing + Exogenous Variables
📈 Price Features
OHLCV + Returns
📊 Technical
RSI, MACD, MA
📰 Sentiment
News Analysis
⏰ Time
T=30 days
3
ML/DL Model
Predict: $\hat{y}_{T+1:T+S}$
LSTM
Architecture
Input
$\mathbf{x}_{1:T}$
Output
$\hat{y}_t$
Confidence
78%
4
Trading Signal
Decision: Buy/Sell/Hold
📈 Signal
BUY
🎯 Target
$155.50
🛡️ Stop
$148.00
📊 Confidence
85%
5
Trading
Action
Action
BUY 100 AAPL
Price
$150.30
Cost
$15,030
Status
Filled
🧠 RL-based Approach
1
State Observation
State: $s_t$
📊 Market
Price, Volume
💼 Portfolio
Positions, Cash
📰 News
Sentiment
⏰ Time
Market Hours
2
Action Space
Actions: $\mathcal{A}$
📈 Buy
Long Position
📉 Sell
Short Position
⏸️ Hold
No Action
💰 Size
Position Size
3
RL Agent
Policy: $\pi_\theta(a_t \mid s_t)$
PPO
Algorithm
State
$s_t$
Action
$a_t$
Reward
$r_t$
4
Reward Function
Reward: $r_t$
📈 Returns
+2.3%
📊 Sharpe
1.8
📉 Drawdown
-5.2%
🎯 Risk
15%
5
Policy Learning
Optimization
Objective
Max $\mathbb{E}[\sum r_t]$
Method
Policy Gradient
Episodes
10,000
Convergence
85%
Key Features
- Real-time Decision Making - Immediate action selection based on market conditions
- Risk Management - Built-in risk controls and position limits
- Multiple Paradigms - Support for both predictive and RL approaches
- Transaction Costs - Realistic simulation of trading costs and slippage
Portfolio Management
Portfolio management focuses on the construction, optimization, and dynamic rebalancing of investment portfolios subject to real-world operational constraints. It supports various objective functions including return maximization, volatility minimization, and Sharpe ratio optimization.
ML/DL-based Approach
Goal: Predict future asset returns or risk estimates
$$\hat{\mathbf{y}}_{T+1:T+S} = \mathcal{F}_\theta(\mathbf{x}_{1:T}, \mathbf{z}_{1:T_{\mathrm{ex}}})$$
Portfolio Weights:
- $\mathbf{w}_{T+1:T+S}$ where $\sum_{i=1}^N w_{t,i} = 1$ and $w_{t,i} \geq 0$
- Determined by optimization procedures (mean-variance, risk parity, etc.)
Where:
- $\hat{\mathbf{y}}_t$ may represent predicted returns, risk scores, or signals
- Weights are optimized based on predictions
RL-based Approach
Sequential Decision Process:
At each time $t$:
- Observe state $s_t$ (constructed from $\mathbf{x}_{1:t}$ and $\mathbf{z}_{1:t}$)
- Select allocation weights $\mathbf{w}_t \in \Delta^{N+1}$
- Receive reward $r_t$ (portfolio return or risk-adjusted reward)
- Transition to next state $s_{t+1}$
Objective: Learn policy $\pi_\theta(\mathbf{w}_t \mid s_t)$ that maximizes:
$$\max_\theta~ \mathbb{E}_{\pi_\theta} \left[\sum_{t=T+1}^{T+S} U(\mathbf{w}_t, \mathbf{x}_t)\right]$$
Where:
- $w_{t,0}$ represents cash position
- $U(\cdot)$ represents cumulative portfolio return
💼 Portfolio Management Workflow
📊 ML/DL-based Approach
1
Multi-Asset Data
Input: $\mathbf{x}_{1:T} \in \mathbb{R}^{N \times T \times D}$
📈 AAPL
$150.25
💻 MSFT
$380.50
🔍 GOOGL
$2,850
📊 VIX
18.5
2
Feature Engineering
Processing + Exogenous Variables
📈 Returns
Historical Returns
📊 Risk
Volatility, VaR
🔗 Correlation
Asset Correlations
⏰ Time
T=30 days
3
ML/DL Model
Predict: $\hat{\mathbf{y}}_{T+1:T+S}$
LSTM
Architecture
Input
$\mathbf{x}_{1:T}$
Output
$\hat{\mathbf{y}}_t$
Assets
N=4
4
Portfolio Weights
Weights: $\mathbf{w}_{T+1:T+S}$
📈 AAPL
35%
💻 MSFT
30%
🔍 GOOGL
25%
💰 Cash
10%
5
Portfolio Management
Action
Expected Return
8.5%
Risk
12.3%
Sharpe Ratio
1.85
Rebalance
Weekly
🧠 RL-based Approach
1
State Observation
State: $s_t$
📊 Market
Prices, Volumes
💼 Portfolio
Current Weights
📰 News
Market Sentiment
⏰ Time
Market Hours
2
Action Space
Actions: $\mathcal{A}$
⚖️ Weights
$\mathbf{w}_t \in \Delta^{N+1}$
💰 Cash
$w_{t,0}$
📈 Assets
$w_{t,1:N}$
🔄 Rebalance
Frequency
3
RL Agent
Policy: $\pi_\theta(\mathbf{w}_t \mid s_t)$
PPO
Algorithm
State
$s_t$
Action
$\mathbf{w}_t$
Reward
$r_t$
4
Reward Function
Reward: $r_t$
📈 Returns
+8.5%
📊 Sharpe
1.85
📉 Drawdown
-5.2%
🎯 Risk
12.3%
5
Policy Learning
Optimization
Objective
Max $\mathbb{E}[\sum U(\mathbf{w}_t, \mathbf{x}_t)]$
Method
Policy Gradient
Episodes
10,000
Convergence
88%
Key Features
- Multi-Asset Allocation - Simultaneous optimization across multiple assets
- Risk Constraints - Position limits, transaction costs, and regulatory constraints
- Dynamic Rebalancing - Adaptive portfolio adjustments over time
- Objective Flexibility - Support for various risk-return objectives
LLM Applications
LLM applications in finance encompass two main categories: general language understanding tasks and sequential decision-making tasks. These applications leverage the power of large language models for financial reasoning, analysis, and decision-making.
Mathematical Formulation
Given: Multi-modal financial inputs
- Unstructured text (news, reports)
- Structured time series (OHLC prices)
- Images (K-line charts)
- Audio/Video (financial broadcasts)
Goal: Train and deploy large language models $\mathcal{M}_\phi$
$$\mathbf{y} = \mathcal{M}_\phi(\mathbf{D}_1, \mathbf{D}_2, ..., \mathbf{D}_K)$$
Where:
- Each $\mathbf{D}_k$ represents an input modality
- $\mathbf{y}$ is the task-specific output
- $\mathcal{M}_\phi$ is the LLM parameterized by $\phi$
Application Categories
- General Language Understanding - Financial text analysis, QA, reasoning
- Sequential Decision-Making - RL training through environment interaction
- Multi-Modal Processing - Integration of text, time series, and visual data
- Tool Utilization - Advanced tool use for financial analysis
🤖 LLM Applications Workflow
📱 Application Types
1
Multi-Modal
Data
Data
Sources
📰 Text
News, Reports
📈 Time Series
OHLC Prices
📊 Images
K-line Charts
🎵 Audio/Video
Financial Broadcasts
2
LLM Processing
Model: $\mathcal{M}_\phi$
🧠 Architecture
Transformer-based
📚 Training
SFT + RL
🎯 Capability
Reasoning
⚡ Output
$\mathbf{y} = \mathcal{M}_\phi(\mathbf{D}_1, ..., \mathbf{D}_K)$
3
Different Tasks
Task Categories
💬 General QA
Financial Questions
📈 Time Series
Forecasting
💼 Trading
BUY/SELL/HOLD
📊 Portfolio
Asset Selection
GRPO Training
Group Relative Policy Optimization (GRPO) is used for efficient LLM training in financial contexts:
- Stage 1: Financial reasoning abilities using financial reasoning datasets
- Stage 2: Real/simulated market environment interaction for practical skills
- Group-level Normalization: Stable learning signals for fine-tuning
- Environment Interaction: Direct trading environment interaction for skill acquisition
🔄 Two-Stage GRPO Training
Stage I: Financial Reasoning Fine-tuning
1
Financial Datasets
Reasoning Tasks
📊 QA
Financial Questions
🧮 Math
Quantitative Analysis
📈 Analysis
Market Scenarios
🎯 Goal
Reasoning Skills
2
RL
Training
Training
Reasoning
📚 SFT
Supervised Learning
🔄 RL
Reinforcement Learning
🎯 Reward
Reasoning Quality
📈 Result
Financial Reasoning
3
Reasoning Capability
Foundation Model
💭 Understanding
Financial Concepts
🧮 Calculation
Quantitative Skills
📊 Analysis
Market Interpretation
🎯 Limitation
No Market Experience
Stage II: Market Environment Learning
4
Market Environment
Real Market Data
📈 Prices
Historical Data
📰 News
Market Events
⚡ Volatility
Market Dynamics
🎯 Goal
Decision Making
5
RL Training
Environment Interaction
🎮 Environment
Market Simulation
🎯 Actions
Trading Decisions
💰 Rewards
Financial Performance
📈 Learning
Trial & Error
6
Deployment Ready
Production Model
💼 Trading
Real Decisions
📊 Analysis
Market Adaptation
🎯 Performance
Financial Outcomes
🔄 Continuous
Learning & Update
Task Integration
FinWorld's modular architecture enables seamless integration across all four task types, allowing researchers and practitioners to:
Cross-Task Learning
- Shared Representations - Common feature extraction across tasks
- Transfer Learning - Knowledge transfer between related tasks
- Multi-Task Training - Joint optimization of multiple objectives
Unified Evaluation
- Standardized Metrics - Consistent evaluation across all tasks
- Comparative Analysis - Direct comparison of different approaches
- Reproducible Results - Standardized protocols for fair comparison
Flexible Deployment
- Pipeline Composition - Easy combination of different task components
- Real-time Processing - Support for live market applications
- Scalable Architecture - From research prototypes to production systems