Never start with data mining. That's where amateurs begin, and it's why they fail. You start with a hypothesis about market inefficiency—a specific reason why you believe you can extract alpha from the market.
Good hypotheses sound like this: "Funding rate extremes lead to mean reversion because excessive positioning creates squeeze conditions." Or "Volume spikes at support levels indicate accumulation and precede bounces." Maybe "On-chain whale accumulation during fear signals smart money positioning."
Bad approaches sound like this: "Let me try different indicator combinations until something works." Or "I'll optimize parameters until the equity curve looks good." The worst: "This YouTube guy's strategy sounds profitable."
Your hypothesis documentation should be thorough and specific. Here's a template that forces you to think clearly:
- **Strategy Name:** Funding Rate Mean Reversion
- **Hypothesis:** When perpetual funding rates reach extreme levels (>0.05% or <-0.02%),
the subsequent 24-48 hours tend to see price move against the crowded side.
- **Market Inefficiency:** Retail leverage positioning tends to overshoot, creating
temporary price deviations that revert as positions are forced to close.
Expected Edge: 55-60% win rate on mean reversion trades with 1.5:1 R:R
Timeframe: 4-hour to daily
Assets: BTC, ETH, and high-liquidity altcoins with perpetual markets
- **Entry Criteria:** Funding rate exceeds 95th percentile (positive) or 5th percentile (negative)
- **Exit Criteria:** Funding normalizes to 25-75th percentile range OR 48-hour time stop OR
fixed stop loss at 3%
Initial Risk: 1% of account per trade
Starting with a hypothesis matters because it limits your degrees of freedom. Fewer parameters to optimize means less overfitting. You understand WHY the strategy works, which provides interpretability when debugging failures. It guides feature selection—you only include relevant data. And when market conditions change, understanding the underlying mechanism helps you adapt.
Data quality determines backtest quality. Garbage in, garbage out—and crypto data is often garbage. You need comprehensive coverage across multiple data types, and each type has specific requirements.
For basic OHLCV data, you need at least 2 years but prefer 5+ years. Sources like Binance, Coinbase, and institutional market data platforms provide reliable feeds. Funding rates need 1 year minimum, 3+ years ideally—get this from derivatives data platforms or exchange APIs. Open interest data requires 1-2 years from derivatives data platforms or on-chain metrics platforms.
On-chain data needs 2-4 years minimum from on-chain metrics platforms or IntoTheBlock. If you're doing microstructure analysis, order book data needs 6 months to 1+ year from institutional market data platforms or Tardis.
Your data cleaning checklist is non-negotiable: no missing periods (gaps must be filled or marked), timestamps in consistent timezone, prices adjusted for any corporate actions, volume in consistent units, no lookahead bias in derived features, and point-in-time accuracy—data available when claimed.
Here's a basic cleaning process that catches most issues:
def clean_ohlcv_data(df):
"""Clean and validate OHLCV data"""
# 1. Check for missing timestamps
expected_periods = pd.date_range(df.index.min(), df.index.max(), freq='1H')
missing = expected_periods.difference(df.index)
print(f"Missing periods: {len(missing)}")
# 2. Fill missing with forward fill (or mark as missing)
df = df.reindex(expected_periods).ffill()
# 3. Remove outliers (prices >10 std from rolling mean)
rolling_mean = df['close'].rolling(24).mean()
rolling_std = df['close'].rolling(24).std()
outliers = abs(df['close'] - rolling_mean) > 10 * rolling_std
df.loc[outliers, 'close'] = rolling_mean[outliers]
# 4. Verify OHLC consistency
assert (df['high'] >= df['low']).all()
assert (df['high'] >= df['open']).all()
assert (df['high'] >= df['close']).all()
# 5. Check volume consistency
assert (df['volume'] >= 0).all()
return df
The train/validation/test split is critical. Never optimize on test data—ever. With 5 years of data from January 2020 to December 2024, use 60% for training (January 2020 to December 2022), 30% for validation (January 2023 to June 2024), and 10% for testing (July 2024 to December 2024).
Develop and tune only on training data. Use validation for hyperparameter selection. Test once at the very end. If you touch test data and iterate, it becomes validation data and you've contaminated your out-of-sample testing.
Strategy implementation separates professionals from amateurs. Poor implementation creates subtle bugs that invalidate your entire backtest. Follow these practices religiously.
Use vectorized operations whenever possible. Loops are slow and error-prone. Instead of iterating through every row checking conditions, use pandas and numpy vectorization for speed and clarity.
Feature generation should be systematic and bug-free:
def generate_features(df):
"""Generate all features needed for strategy"""
# Price features
df['return_1h'] = df['close'].pct_change(1)
df['return_24h'] = df['close'].pct_change(24)
df['price_vs_sma20'] = df['close'] / df['close'].rolling(20).mean()
# Volatility
df['atr'] = calculate_atr(df, period=14)
df['volatility'] = df['return_1h'].rolling(24).std() * np.sqrt(8760)
# Technical indicators
df['rsi'] = calculate_rsi(df['close'], period=14)
df['macd'], df['macd_signal'], df['macd_hist'] = calculate_macd(df['close'])
# Funding rate features (if available)
if 'funding_rate' in df.columns:
df['funding_percentile'] = df['funding_rate'].rolling(720).rank(pct=True)
return df
Signal generation should be clear and logical:
def generate_signals(df):
"""Generate entry and exit signals"""
# Entry signals
df['long_signal'] = (
(df['funding_percentile'] < 0.05) & # Extreme negative funding
(df['rsi'] < 35) & # Oversold
(df['price_vs_sma20'] < 0.98) # Below MA
).astype(int)
df['short_signal'] = (
(df['funding_percentile'] > 0.95) & # Extreme positive funding
(df['rsi'] > 65) & # Overbought
(df['price_vs_sma20'] > 1.02) # Above MA
).astype(int)
# Exit signals
df['exit_long'] = (
(df['funding_percentile'] > 0.5) | # Funding normalized
(df['rsi'] > 60) # Momentum exhausted
)
df['exit_short'] = (
(df['funding_percentile'] < 0.5) |
(df['rsi'] < 40)
)
return df
Position management requires careful tracking of entry prices, stop losses, and profit targets:
class Position:
def __init__(self, entry_price, direction, size, entry_time):
self.entry_price = entry_price
self.direction = direction # 1 for long, -1 for short
self.size = size
self.entry_time = entry_time
self.stop_loss = self._calculate_stop()
self.take_profit = self._calculate_target()
def _calculate_stop(self):
if self.direction == 1:
return self.entry_price * 0.97 # 3% stop for longs
else:
return self.entry_price * 1.03 # 3% stop for shorts
def _calculate_target(self):
if self.direction == 1:
return self.entry_price * 1.045 # 4.5% target for longs
else:
return self.entry_price * 0.955
def check_exit(self, current_price, current_time):
"""Check if position should be closed"""
# Stop loss
if self.direction == 1 and current_price <= self.stop_loss:
return 'stop_loss'
if self.direction == -1 and current_price >= self.stop_loss:
return 'stop_loss'
# Take profit
if self.direction == 1 and current_price >= self.take_profit:
return 'take_profit'
if self.direction == -1 and current_price <= self.take_profit:
return 'take_profit'
# Time stop (48 hours max)
if (current_time - self.entry_time).total_seconds() > 48 * 3600:
return 'time_stop'
return None
Most backtests fail because they assume perfect execution in a frictionless world. Reality is messier. Transaction costs, slippage, and execution delays kill strategies that look profitable on paper.
Here's what typical costs look like: maker fees range from 0.01% to 0.04% (use 0.02%), taker fees from 0.04% to 0.10% (use 0.06%), spreads from 0.01% to 0.05% (use 0.02%), and slippage from 0.02% to 0.20% depending on order size (use 0.05% as baseline). Total round-trip costs typically range from 0.08% to 0.40%—I recommend assuming 0.15% for most strategies.
Slippage modeling requires sophistication beyond simple percentages:
def estimate_slippage(order_size_usd, avg_daily_volume, volatility):
"""
Estimate slippage based on order size and market conditions
Uses square-root impact model: slippage ∝ sqrt(size / volume)
"""
size_ratio = order_size_usd / (avg_daily_volume * 0.01) # % of daily volume
base_slippage = 0.0001 * np.sqrt(size_ratio) # Square-root impact
volatility_factor = 1 + volatility # Higher volatility = more slippage
return base_slippage * volatility_factor
Your execution assumptions matter enormously. Naive backtests assume you get filled at close prices with 100% fill rate and instant execution. Realistic backtests add slippage to close prices, assume 95% fill rate for limit orders (100% for market orders), include 1-5 second execution latency, and model partial fills for large orders.
Capital management prevents you from taking impossible positions:
class Portfolio:
def __init__(self, initial_capital=10000, max_position_pct=0.2, max_leverage=3):
self.initial_capital = initial_capital
self.capital = initial_capital
self.max_position_pct = max_position_pct
self.max_leverage = max_leverage
self.positions = []
self.trade_history = []
def calculate_position_size(self, entry_price, stop_loss, risk_pct=0.01):
"""[calculate position size](/tools/position-size-calculator) based on risk"""
risk_amount = self.capital * risk_pct
price_risk = abs(entry_price - stop_loss) / entry_price
# Position size to risk specified amount
position_size = risk_amount / price_risk
# Cap at max position percentage
max_size = self.capital * self.max_position_pct
position_size = min(position_size, max_size)
# Cap at max leverage
max_leveraged = self.capital * self.max_leverage
position_size = min(position_size, max_leveraged)
return position_size
Your backtest engine needs to handle the complexity of multi-position strategies, realistic execution, and comprehensive logging:
def backtest_strategy(df, portfolio, strategy_params):
"""
Run backtest over historical data
"""
results = {
'trades': [],
'equity_curve': [],
'daily_returns': []
}
for i in range(strategy_params['lookback'], len(df)):
current_bar = df.iloc[i]
current_time = df.index[i]
# Update existing positions
for position in portfolio.positions[:]: # Copy for safe iteration
exit_reason = position.check_exit(current_bar['close'], current_time)
if exit_reason:
trade = close_position(portfolio, position, current_bar, exit_reason)
results['trades'].append(trade)
# Check for new entry signals
if len(portfolio.positions) < strategy_params['max_positions']:
if current_bar['long_signal']:
open_position(portfolio, 'long', current_bar, strategy_params)
elif current_bar['short_signal']:
open_position(portfolio, 'short', current_bar, strategy_params)
# Record equity
equity = calculate_equity(portfolio, current_bar['close'])
results['equity_curve'].append({
'timestamp': current_time,
'equity': equity
})
return results
Once you have results, analyze them systematically:
def analyze_results(results, initial_capital):
"""Calculate performance metrics"""
trades = pd.
DataFrame(results['trades'])
equity = pd.
DataFrame(results['equity_curve'])
metrics = {
'total_return': (equity['equity'].iloc[-1] / initial_capital - 1) * 100,
'total_trades': len(trades),
'win_rate': (trades['pnl'] > 0).mean() * 100,
'avg_win': trades[trades['pnl'] > 0]['pnl'].mean(),
'avg_loss': trades[trades['pnl'] < 0]['pnl'].mean(),
'profit_factor': trades[trades['pnl'] > 0]['pnl'].sum() / abs(trades[trades['pnl'] < 0]['pnl'].sum()),
'max_drawdown': calculate_max_drawdown(equity['equity']),
'sharpe_ratio': calculate_sharpe(equity['equity']),
'calmar_ratio': calculate_calmar(equity['equity'])
}
return metrics
Before proceeding to validation, run sanity checks. You need more than 50 trades for statistical significance. Win rates should fall between 30-70%—extreme values are suspicious. Trade frequency should match expectations based on your signal frequency. The equity curve shouldn't show unrealistic jumps. Position sizing should never exceed your maximum limits.
Walk-forward analysis tests whether your optimization generalizes to unseen data. This is where most overfitted strategies die. If your strategy can't maintain performance when optimized on one period and tested on the next, it won't work in live trading.
The process involves sliding windows: train on 12 months, test on the next 3 months, then advance by 3 months and repeat. Each iteration optimizes parameters on the training window, then tests those optimal parameters on the out-of-sample test window.
def walk_forward_analysis(df, strategy, train_period=12, test_period=3):
"""
Perform walk-forward optimization
train_period: months to train on
test_period: months to test on
"""
results = []
for start_idx in range(0, len(df) - train_period - test_period, test_period):
# Define windows
train_start = start_idx
train_end = start_idx + train_period
test_start = train_end
test_end = test_start + test_period
# Train (optimize) on training window
train_data = df.iloc[train_start:train_end]
optimal_params = optimize_strategy(train_data, strategy)
# Test on out-of-sample window
test_data = df.iloc[test_start:test_end]
test_results = backtest_strategy(test_data, strategy, optimal_params)
results.append({
'train_period': (df.index[train_start], df.index[train_end]),
'test_period': (df.index[test_start], df.index[test_end]),
'optimal_params': optimal_params,
'in_sample_sharpe': calculate_sharpe(train_results),
'out_sample_sharpe': calculate_sharpe(test_results),
'test_return': test_results['total_return']
})
return results
Walk-Forward Efficiency (WFE) measures how much out-of-sample performance degrades compared to in-sample optimization:
def calculate_wfe(walk_forward_results):
"""
Walk-Forward Efficiency (WFE)
WFE = Out-of-Sample Performance / In-Sample Performance
Target: >50% indicates robustness
"""
in_sample_avg = np.mean([r['in_sample_sharpe'] for r in walk_forward_results])
out_sample_avg = np.mean([r['out_sample_sharpe'] for r in walk_forward_results])
wfe = (out_sample_avg / in_sample_avg) * 100
return wfe
WFE above 70% indicates excellent robustness—proceed to Monte Carlo testing. 50-70% shows good robustness but review for minor overfitting. 30-50% raises significant concerns about overfitting. Below 30% indicates poor robustness and requires major strategy redesign.
Different market conditions require different strategies. A mean reversion strategy that works in sideways markets might get destroyed in strong trends. Test each regime separately to understand where your strategy has edge and where it doesn't.
Regime classification combines trend and volatility analysis:
def classify_regimes(df):
"""
Classify market regimes based on trend and volatility
"""
# Trend classification
df['sma50'] = df['close'].rolling(50).mean()
df['sma200'] = df['close'].rolling(200).mean()
df['trend'] = 'neutral'
df.loc[(df['close'] > df['sma50']) & (df['sma50'] > df['sma200']), 'trend'] = 'bullish'
df.loc[(df['close'] < df['sma50']) & (df['sma50'] < df['sma200']), 'trend'] = 'bearish'
# Volatility classification
df['realized_vol'] = df['close'].pct_change().rolling(24).std() * np.sqrt(8760)
vol_percentile = df['realized_vol'].rank(pct=True)
df['volatility'] = 'normal'
df.loc[vol_percentile > 0.75, 'volatility'] = 'high'
df.loc[vol_percentile < 0.25, 'volatility'] = 'low'
# Combined regime
df['regime'] = df['trend'] + '_' + df['volatility']
return df
Run separate backtests for each regime:
def backtest_by_regime(df, strategy):
"""Test strategy separately in each regime"""
regime_results = {}
for regime in df['regime'].unique():
regime_data = df[df['regime'] == regime]
if len(regime_data) > 100: # Minimum data for significance
results = backtest_strategy(regime_data, strategy)
regime_results[regime] = analyze_results(results)
return regime_results
Understanding regime performance lets you deploy capital intelligently. Maybe your strategy crushes it during bullish_normal and neutral_normal conditions (deploy fully), shows modest edge during bullish_high and neutral_low (reduce position size), and struggles during bearish regimes (skip entirely or reverse signals).
This regime awareness prevents you from deploying capital blindly when conditions don't favor your strategy's underlying logic.
Monte Carlo simulation reveals the range of possible outcomes beyond your single historical path. Markets could have unfolded differently with the same underlying dynamics, and Monte Carlo shows you how your strategy would have performed across those alternative histories.
Trade shuffling tests sequence dependency. If your strategy relies on trades occurring in a specific order, shuffling reveals this fragility:
def monte_carlo_shuffle(trades, num_simulations=10000):
"""
Shuffle trade sequence and recalculate equity curves
"""
results = []
for _ in range(num_simulations):
shuffled = trades.sample(frac=1, replace=False)
equity_curve = calculate_equity_from_trades(shuffled)
results.append({
'final_equity': equity_curve[-1],
'max_drawdown': calculate_max_drawdown(equity_curve),
'sharpe': calculate_sharpe(equity_curve)
})
return pd.
DataFrame(results)
Bootstrap resampling samples trades with replacement to estimate confidence intervals:
def monte_carlo_bootstrap(trades, num_simulations=10000):
"""
Bootstrap sample trades to estimate confidence intervals
"""
results = []
for _ in range(num_simulations):
sampled = trades.sample(frac=1, replace=True)
metrics = calculate_metrics(sampled)
results.append(metrics)
return pd.
DataFrame(results)
Analyzing Monte Carlo results gives you confidence intervals:
def analyze_monte_carlo(mc_results):
"""Calculate confidence intervals from Monte Carlo"""
analysis = {
'final_equity': {
'median': mc_results['final_equity'].median(),
'p5': mc_results['final_equity'].quantile(0.05),
'p95': mc_results['final_equity'].quantile(0.95)
},
'max_drawdown': {
'median': mc_results['max_drawdown'].median(),
'p5': mc_results['max_drawdown'].quantile(0.05),
'p95': mc_results['max_drawdown'].quantile(0.95)
},
'sharpe': {
'median': mc_results['sharpe'].median(),
'p5': mc_results['sharpe'].quantile(0.05),
'p95': mc_results['sharpe'].quantile(0.95)
},
'probability_profitable': (mc_results['final_equity'] > 10000).mean()
}
return analysis
Red flags in Monte Carlo results include the 5th percentile showing capital loss, your original backtest significantly above the 95th percentile (suggesting you got lucky), or a wide spread between 5th and 95th percentiles (indicating high uncertainty). Good strategies show at least 80-90% probability of profitability across simulations.
Statistical significance ensures your results exceed what random chance would produce. Too many traders deploy strategies based on noise, not signal.
Test if your returns are statistically different from zero:
from scipy import stats
def test_significance(strategy_returns, benchmark_returns=None):
"""
Test if strategy returns are statistically significant
"""
# One-sample t-test against zero
t_stat, p_value = stats.ttest_1samp(strategy_returns, 0)
result = {
'mean_return': strategy_returns.mean(),
'std_return': strategy_returns.std(),
't_statistic': t_stat,
'p_value': p_value,
'significant_5pct': p_value < 0.05,
'significant_1pct': p_value < 0.01
}
if benchmark_returns is not None:
# Two-sample t-test against benchmark
t_stat_bench, p_value_bench = stats.ttest_ind(
strategy_returns, benchmark_returns
)
result['vs_benchmark_t'] = t_stat_bench
result['vs_benchmark_p'] = p_value_bench
return result
The minimum sample sizes required for statistical significance depend on your Sharpe ratio. A strategy with 0.5 Sharpe needs 16 years for 95% significance. 1.0 Sharpe needs 4 years. 1.5 Sharpe needs 2 years. 2.0 Sharpe needs 1 year. 3.0+ Sharpe needs only 6 months.
When testing multiple strategies or parameters, you must correct for multiple hypothesis testing:
def bonferroni_correction(p_values, alpha=0.05):
"""
Adjust for multiple hypothesis testing
"""
n_tests = len(p_values)
adjusted_alpha = alpha / n_tests
significant = [p < adjusted_alpha for p in p_values]
return adjusted_alpha, significant
If you test 20 parameter combinations, your significance threshold drops from 0.05 to 0.0025. Only p-values below this adjusted threshold represent genuine significance rather than luck.
Before risking real capital, validate your strategy in live market conditions with paper trading. This catches issues that backtesting can't reveal—data quality problems, execution timing, latency effects, and market microstructure changes.
Your paper trading checklist must validate that signals fire when expected, fills occur at expected prices, latency remains acceptable from signal to execution, position sizing calculations work correctly, risk management triggers properly, and data feeds remain clean without gaps or errors.
Build a paper trading framework that mirrors your live execution:
class PaperTrader:
def __init__(self, strategy, initial_capital=10000):
self.strategy = strategy
self.capital = initial_capital
self.positions = []
self.trade_log = []
def run_live_signal_check(self, current_data):
"""Check for signals on live data"""
signal = self.strategy.generate_signal(current_data)
if signal:
self.log_signal(signal, current_data)
# Simulate fill at next available price
fill_price = self.estimate_fill_price(signal, current_data)
self.execute_paper_trade(signal, fill_price)
def reconcile_with_backtest(self, backtest_results):
"""Compare paper trading to backtest expectations"""
paper_metrics = self.calculate_metrics()
comparison = {
'win_rate_diff': paper_metrics['win_rate'] - backtest_results['win_rate'],
'avg_return_diff': paper_metrics['avg_return'] - backtest_results['avg_return'],
'trade_freq_diff': paper_metrics['trades_per_month'] - backtest_results['trades_per_month']
}
# Flag significant deviations
if abs(comparison['win_rate_diff']) > 10:
print("WARNING: Win rate significantly different from backtest")
return comparison
Paper trading duration depends on your strategy frequency. High-frequency strategies need 2 weeks minimum with 100+ trades. Day trading strategies need 1 month with 50+ trades. Swing trading requires 3 months with 30+ trades. Position trading needs 6 months with 20+ trades.
Common issues include lookahead bias in your backtest (signals arrive earlier in paper trading), overly optimistic fills (worse execution in paper trading), data source differences (different signals), and weekend/holiday gaps causing unexpected behavior.
Step 1: Define Hypothesis
└── Clear market inefficiency thesis
Step 2: Gather Data
└── Clean, validated historical data
Step 3: Implement Strategy
└── Vectorized, realistic logic
Step 4: Configure Simulation
└── Transaction costs, slippage, capital
Step 5: Initial Backtest
└── Run on training data, sanity check
Step 6: Walk-Forward Analysis
└── WFE > 50% required
Step 7: Regime Testing
└── Understand where strategy works/fails
Step 8: Monte Carlo
└── 90%+ simulations profitable
Step 9: Statistical Testing
└── p-value < 0.05 (Bonferroni adjusted)
Step 10: Paper Trading
└── Validate in live conditions
Only proceed to live trading if ALL steps pass!
You need a minimum of 50 trades for basic validity, but 100+ is preferred. For strategies with extreme win rates (below 30% or above 70%), you need even more trades. The rule of thumb is simple—you need enough trades that adding 10 more wouldn't significantly change your key metrics.
Absolutely, and this isn't optional. Include maker/taker fees, spread, and estimated slippage. A strategy that's profitable before costs often becomes unprofitable after realistic friction. For high-frequency strategies, costs can consume 50% or more of gross profits, turning apparent winners into guaranteed losers.
Key warning signs include walk-forward efficiency below 50%, performance that collapses with small parameter changes, results that are dramatically better than simple benchmarks, single-point optimal parameters with cliff-like drop-offs nearby, and in-sample performance that significantly exceeds out-of-sample performance.
You need 2 years minimum covering both bull and bear markets, but 4-5 years is ideal for statistical significance and regime diversity. Very old data from before 2020 may not reflect current market dynamics, so weight recent data more heavily in your analysis.
Yes, but ensure you're testing on data the AI didn't train on. If the AI was trained on 2020-2023 data, only backtest its signals on 2024 data. Otherwise, you're testing in-sample performance, which tells you nothing about future performance.
Quarterly reviews are reasonable for most strategies. Re-run validation when performance degrades significantly, market regimes change, you modify parameters, or new data reveals previously unseen conditions that could affect your strategy's performance.
Rigorous backtesting transforms strategy development from gambling to engineering. The difference between successful traders and the majority who fail comes down to following the key steps that actually matter.
Start with hypothesis-first development—begin with a clear market inefficiency thesis, not data mining expeditions. Ensure data quality with clean, gap-free data that maintains point-in-time accuracy to prevent lookahead bias. Configure realistic simulation that includes transaction costs, slippage, and capital constraints rather than assuming perfect execution.
Achieve walk-forward validation with greater than 50% walk-forward efficiency before moving to the next step. Develop regime awareness by understanding where your strategy works and where it fails. Run Monte Carlo analysis to ensure 90% or more of simulations remain profitable. Confirm statistical significance so your results exceed random chance with appropriate multiple testing corrections. Finally, validate through paper trading in live conditions before deploying any capital.
The traders who follow this rigorous process deploy strategies with confidence. Those who skip steps deploy hope, and hope isn't a trading strategy.
Thrive provides the tools you need for rigorous strategy validation. You get years of clean price, volume, and on-chain data for comprehensive backtesting. Test AI signals across historical conditions with proper regime analysis. Get comprehensive analytics for strategy evaluation with automatic regime classification.
Validate signals in live conditions risk-free through paper trading, then track strategy performance with real capital through live monitoring. From hypothesis to profitable strategy, we've got you covered.
→ Start Backtesting with Thrive
